Wiring the Brain

What is cognition?

noreply@blogger.com (Kevin Mitchell) — Sat, 08 Mar 2025 12:53:00 +0000

“Cognition” is one of those terms that we use a lot, but which is actually rather hard to define. The question of what constitutes cognition looms large in debates about whether other organisms, besides humans, do it or have it. Most people are willing to grant that many animals are in some way cognitive, but what about simpler critters, like nematodes or jellyfish? Are they capable of cognition? For that matter, what about bacteria or plants? Can they cognize? How about artificial systems, like large language models?

Does cognition require a nervous system? Does it necessarily involve “thinking”? Or some kind of “mental” activity? Once you start defining it in that rather circular fashion, you quickly realise we don’t have good definitions of “thinking” or “mental” either!

Is cognition just “what cognitive scientists study”? Or are they only studying a particular version of it – the kind of cognition we see in humans? Maybe there is a less anthropocentric way to define it – one that recognizes some abstract principles that may be realised in different ways in different kinds of organisms.

To that end, I posed the question on Bluesky: “how would you define ‘cognition’”? There were lots of replies, from neuroscientists, cognitive scientists, psychologists, philosophers, computer scientists, and others. (Thread here, from January 2025)

I’m hugely grateful to everyone who replied and hope they won’t mind me listing their (anonymised) answers below. What’s fascinating is how diverse the replies were. They run the gamut from “information processing by goal-directed systems” to “all the processes between perception and action” to definitions that use high-level human psychological terms like “thinking”, “planning”, and other kinds of “mental operations”. Many definitions emphasized what cognition is for: problem-solving and using information to adapt to the world.

You can peruse the list below. I’ll give my own definition at the bottom, along with some further thoughts on how – beyond simply looking for a definition – we might usefully think about cognition. Here’s the list:

Cognition is:

The stuff that happens between 'perception' and 'action'

Anything about how organisms interact with the world that is not sensory or motor.

Considering most of the so-called cognition happens unconsciously and our poor knowledge of the brain functioning, cognition is currently a black box between sensory stimuli and motor actions.

Information processing in the brain.

Information processing done by goal-directed systems.

The problem-solving part of the organism, excluding management of bodily functions.

The information processing aspect of living systems. So it’s intimately entwined with life, but not the same.

Processes that allow you to act efficiently in contexts with different degrees of uncertainty.

The set of things needed to support skillful and intentional activity in the world

How our minds work to interpret the world

Mental actions or processes. Neural actions or processes may be the underlying mechanism, but are not themselves called "cognitive." Cognitive neuroscience studies the biological mechanisms of cognition.

Any process where the input, output, or intermediate steps (preferably all) is mental/phenomenal/in the mind.

All things providing knowledge/meaning of the world.

Reasoning, decision-making, planning.

The set of mental processes involved in acquiring, processing, storing, and using information

The ability to synthesize information contained within external stimuli into meaningful interpretations of one's surroundings and current state.

Sensory and other information-processing mechanisms an organism has for becoming familiar with, valuing, and interacting productively with features of its environment in order to meet existential needs.

A property of an agent whereby the behaviour is a kind of possibly partial function/algorithm of some sensory input, over the agent's history.

Brain's specific active contribution to mind is what we call cognition.

The ability of an animal to separate itself from the moment in which it is living and to contemplate the past, predict the future, and act accordingly

Some task that requires non-trivial amounts of information processing, pattern recognition, working with levels of abstraction and composing of solutions to other cognitive or information processing tasks.

The entirety of mental process of acquiring knowledge and understanding through thought, experience, and the senses and using this culmination of knowledge to guide future behaviors.

Body-mind categories of understanding that help an organism interpret, generalise, schematise, and systematise complex sets of sense data into relevant information in its milieu.

What enables an organism to adapt to its environment at a speed faster than what's achieved by evolutionary changes to its body.

Creating and maintaining an internal mental model of the outside world.

Non-equilibrium thermodynamics sustained by information processing.

The generation of meaning in relation to the environment. It is enacting a world.

A concept which exists in the mind of human observers, variously defined and applied to conceptually delineate some subset of adaptive activities in living systems... ...and which is often mistaken for a naturally delineated constituent property of the living systems under study.

And, finally, one respondent pointed out this very helpful resource, where someone did a very similar exercise, with equally diverse answers: http://www.vernon.eu/euCognition/definitions.htm

For what it’s worth, here’s my own definition:

Cognition is using information to solve problems and guide adaptive behavior.

To be very clear, I’m not claiming that’s the “right” definition! I don't think there is a right definition. It’s a pragmatic issue – a question of what kind of thing we are aiming to pick out in the world, and why. What are we trying to achieve in recognising and naming this capacity or type of process? For me, the goal is to get at the underlying principles at play that make some activity cognitive.

For that reason, I tried to couch my definition as abstractly as possible. I think it is broad enough to encompass capacities that share underlying principles but that vary hugely in sophistication across different kinds of creatures.

I would personally argue that even simple phenomena like bacteria integrating information about what is out in the world and mounting an appropriate response to the situation as a whole entails at least a basal kind of cognition (and agency). It’s not a full-blown psychological process, and there’s no reason to think of it as mental (whatever that means), but it’s more than a simple stimulus-response mechanism.

That position is clearly arguable, of course, and there’s not really much point getting hung up on semantics. No one needs to be the “cognition” police. What’s more interesting is to try and dig in a bit – to move beyond that simple question to try and tease out what the concept of cognition is actually getting at.

My original question was simply: How would you define cognition?

I could have asked that in another way, which might have helped avoid the more circular, anthropocentric definitions: How would you define cognition without appealing to the terms “mind” or “mental” or words like “thinking”? That exercise might force some deeper introspection of the underlying principles.

But there are other questions we could ask that might get at something deeper, beyond the mere definition of the term:

What is cognition for? What kinds of tasks are “cognition-hungry”? I mean, what sorts of problems require cognition and what sorts don’t? Getting energy through photosynthesis doesn’t, but hunting prey does. Digesting food doesn’t, but remembering where you cached it does. Walking on an even surface doesn’t, but climbing over a wall does. Right? (That last one is certainly arguable).

We could also ask, from a third-person perspective, what kinds of behaviors would we take as evidence that “cognition” is happening or that a system is “cognitive”? Or, zooming out, what kinds of lifestyles, ecologically speaking, would favour the evolution of cognitive systems?

Here is a rough list of things we might consider as elements or functions of cognition, arranged in ascending order of sophistication:

· Behaving adaptively

· Using information to behave adaptively

· Integrating or synthesising information to behave adaptively and flexibly

· Operating over information to solve problems in order to behave adaptively and flexibly

· Interpreting information in the context of stored knowledge

· Operating over semantic information at a symbolic level

· Operating over symbolic information at a conscious level

It would take some work to flesh out the differences between “using”, “integrating”, “synthesising”, “interpreting”, and “operating over” in those contexts. I’m not even sure what I mean by all those terms, except that the ascending scale loosely implies a greater degree of internal goings-on, more linked to stored schematic knowledge and higher-order, maybe even abstract concepts. So the “information” in question may be simply incoming streams of sensory data, but could also refer to semantically meaningful information, even possibly conceptual information that is divorced from current stimuli, or encoded at a symbolic level. Where we might want to draw the “cognitive” line is thus very open to debate.

Perhaps we could think of adaptive processes in ascending order too:

· Control

· Computation

· Cognition

· Cogitation

I certainly don’t think that all cognition has to be conscious or involve what we think of in humans as cogitation – operations in some kind of mentalese. In fact, I think it’s worth asking whether cognition necessarily implies mentality at all. Or is doing cognitive processes mentally just one of the ways to do them? Is a computer that solves some problems algorithmically doing cognition?

The most basic definitions in the list above almost seem co-extensive with “life”. Is all life cognitive? Humberto Maturana and Francisco Varela influentially argued so. And some contemporary scientists like Mike Levin argue that even processes of embryonic development are essentially cognitive.

Does the word lose its usefulness if we extend it so far? While cognition itself refers to some internal processes, we typically recognise it in the capacities of organisms to solve problems out in the world. Is a bacterium regulating its profile of gene expression in response to information about changing circumstances engaging in a cognitive process? Or does the term “physiology” do the trick? Are the cells in an embryo doing cognition to figure out where they are and what they should turn into? Or is “development” already sufficient to capture those processes?

Perhaps we should say that cognition is something that a whole organism does. There may be all kinds of adaptive, responsive processes happening within an organism that might not rise to this level. Again, it depends on what we want the word to pick out. All living things are entities that use information to solve the most basic problem of persisting. But are they all using it in a cognitive way?

I don't have any answers, but perhaps the questions themselves are useful.

Eugenics, statistical hubris, and unknowable unknowns in human genetics

noreply@blogger.com (Kevin Mitchell) — Mon, 13 Jan 2025 11:27:00 +0000

A new paper just out in Nature, by Peter Visscher and colleagues (including bio-ethicist Julian Salvulescu) explores the idea of polygenic genome editing of human embryos to reduce the risk of common diseases. This is, to say the least, a controversial idea, and a decidedly fantastical one. The authors present the results of statistical modelling which suggests that editing a small number of risk variants in each embryo’s genome could dramatically reduce the risk of a number of common disorders. But there are good reasons (lots of them) to doubt the assumptions on which this modelling is based and to have serious concerns about possibly deleterious unintended consequences of such interventions.

I co-wrote a commentary on the article with geneticist Shai Carmi and law professor and bio-ethicist Hank Greely, outlining some of the limitations of the modelling and our concerns over the dangers of the proposed approach. I’ll expand on those points here.

First, the development of CRISPR genome-editing technologies offers the promise of precise intervention in people’s genetic material. Where they carry some mutation that causes disease, it may be possible to use CRISPR tools to edit that mutation back to its “wild-type” sequence or to intervene in some other way that ameliorates the condition. This technique is already being used to edit the genome in some cells of people suffering from known “monogenic” conditions – ones where the disease is caused by a single, well-defined and well-studied mutation – such as sickle-cell disease. Importantly, the cells that are genetically modified in this case are blood-producing stem cells (taken out of the patient, modified, and put back in), not cells in the germline. As a result, the edits to the genome will not be passed on to offspring of these patients.

The idea that you could apply this in an embryo known to carry disease-causing mutations (in an in vitro fertilisation scenario) is an obvious extension of this approach, but one that is fraught with ethical issues. This would “cure” the resultant individual in the sense of removing or reversing the disease-causing mutations from all their cells. But it would also mean that they would pass on whatever edits are made to their offspring. That’s a big step – it’s not just treating a single patient, it’s affecting subsequent generations, and, depending on how widely it is deployed, could change the gene pool of the population at large.

Now, you might argue that, for mutations like those causing sickle-cell anaemia or cystic fibrosis or Huntington’s disease, that it would just be a good thing if these mutations were indeed removed from the population, or at least reduced in frequency. But good for whom? Who has a legitimate interest in the genetic health of individuals or the population at large? Is this just a question of whether parents should have the right to edit the genomes of their prospective children? Or should we also be concerned with whether the state should actively support (or possibly even mandate) such interventions? When does individual medicine become public health? And when does public health become eugenics? These questions crop up again below.

One immediate concern in these discussions is of course whether the CRISPR methods being used are actually: (i) effective, and (ii) safe. That is, can we be sure that they will really succeed in making the desired edits? And can we also be sure that the approach won’t lead to any “off-target effects” or other unpredictable changes elsewhere in the genome? These are very open questions and many researchers are actively working to increase the precision and reliability of CRISPR methods. So, for the sake of argument, let’s assume we really can make just the genomic edits we want.

The next question is whether the edits themselves will necessarily have (only) beneficial effects. Answering this depends on how confident we are in our current understanding of the relationships between genotypes and phenotypes. For rare mutations causing serious diseases, this seems fairly straightforward. However, even there, there is evidence for some conditions – like sickle-cell anaemia itself, for example – that some mutations that cause disease when present in two copies, have a benefit when present in only one copy – in this case providing some protection against malaria. That said, such situations seem to be very rare and most disease-causing mutations are frankly and uncomplicatedly deleterious – it’s just bad to have them.

But there are other situations where the waters get considerably muddier. This is exemplified by the discovery that very rare mutations of the gene CCR5, which encodes a chemokine receptor involved in immune functions, confer resistance to infection by HIV. (This is because CCR5 is actually the “receptor” for the HIV virus – the means by which the virus attaches itself to immune cells to enable infection). Here is a rare mutation that seems to have beneficial effects – potentially life-saving, in fact.

This is the gene that was infamously targeted by disgraced researcher He Jiankui, who, in 2018, announced that he had edited the genomes of human embryos in order to introduce mutations in the CCR5 gene, and that two babies with such manipulated genomes had been born (discussed in detail in this book by Hank Greely). This intervention was widely condemned, for many reasons, including the fact that informed consent had not been obtained, and that the safety and efficacy of the intervention was entirely unproven and, in essence, untested. It was simply experimenting on human beings. He Jiankui was subsequently jailed although he has since been released and appears unrepentant and intent on further similar efforts.

Apart from the obvious ethical issues, there were two major scientific problems with his “experiments”: first, the desired edits were not in fact made precisely, and it is not known whether other off-target effects may also have occurred. And second, even if those precise edits had been made, there is no guarantee that they would only have had positive effects. The CCR5 protein does not just sit around in our cells waiting to be a receptor for a virus that will kill us. It has normal functions, not just in the immune system, but also unexpected ones, for example in the regulation of bone physiology and even in synaptic plasticity and learning and memory. Knocking it out is thus highly likely to have deleterious effects on other functions – what geneticists call “pleiotropic” effects (from the Greek words pleio, which means "many," and tropic, which means "affecting.")

This case thus illustrates a crucial point: rare mutations that are “protective” for one disease (i.e., statistically associated with reduced risk) are quite likely to have negative pleiotropic effects on other traits. That is, they may be rare for a reason (even if we don’t know what that is). This point is important for the current discussion, because it is precisely these kinds of rare protective mutations that Visscher and colleagues propose to introduce into human embryos.

Targeting common disorders

The target of Visscher and colleagues is not rare monogenic conditions or resistance to infectious agents like viruses. They are concerned, rather, with common disorders, specifically including: coronary artery disease, Alzheimer’s disease, major depressive disorder, diabetes, and schizophrenia. These conditions, which collectively affect millions of people worldwide, are substantially, but not completely heritable. That is, a substantial proportion of the variance in risk for these conditions across the population is due to genetic variation. (Roughly speaking, risk for these conditions loosely “runs in the family”).

But unlike conditions like cystic fibrosis or sickle-cell anaemia, these conditions are not monogenic – they are not caused by mutations in one specific gene. They are polygenic – that is, the genetic risk is due to the combined effects of thousands (or even millions) of genetic variants across the genome. Over the past couple of decades, researchers have used a technique called genome-wide association studies (or GWAS) to identify specific genetic variants (known as single-nucleotide polymorphisms or SNPs) associated with risk for these various conditions.

Each SNP is a site in the genome where the DNA sequence comes in a couple of versions (or “alleles”) across the population – it might be a “C” in some people and an “A” in others, for example. It turns out there are millions of such polymorphic sites in the human genome. If we want to look for variants associated with disease, we simply have to ask whether the frequency of one form (say the “C” version) is higher in people with the condition than without. (That’s exactly analogous to asking whether the frequency of smoking is higher in people with lung cancer than without).

These efforts have been highly successful. Thousands of such “risk variants” have been discovered for the common disorders mentioned above, and for many others. Each such variant typically has a tiny statistical effect on risk, when considered singly (which is why it typically requires samples of tens or even hundreds of thousands of people to detect the tiny differences in frequency between cases and controls). But considered collectively, they can account for a substantial proportion of the total genetic variance in risk.

Now, if that’s the situation, you might be wondering how genome editing could possibly ever be profitably applied to these conditions. If each person’s genetic risk is due to thousands of such SNPs, then how could editing only a few of them make any appreciable difference? That is exactly what Visscher and colleagues explore and they present results of statistical modelling that argue that editing even 5-10 risk variants could have substantial effects on risk – huge effects, in fact! (This is why the paper is in Nature 😉).

They estimate that editing a set of 10 of the variants with largest individual effects (and highest frequency in the population) could reduce the risk of common diseases by anywhere from 3-fold to 60-fold, depending on the condition!

The fact that they selected risk variants that were at very high frequency in the population is important. It might seem odd that there could be sites in the genome where by far the most common variant (at frequencies of 80-99%) is a risk variant for a disorder. That is almost like saying having a typical genome increases your risk of these conditions, which seems a bit weird. But if you look at it from the opposite direction, you can interpret it instead as evidence that the rare allele at each of those sites (the one that only 1-20% of people carry) is actually protective against the disorder. Those are just two different ways of expressing the same statistical observation of a frequency difference between cases and controls.

Focusing on those relatively rare protective variants is a promising approach, as most people will carry the common versions at most of the 10 sites chosen. (Of course, since we actually have two copies of each chromosome, there are really 20 sites at issue). That means that those sites represent a fairly generic target for editing, which is predicted to be applicable to most people in the population. So, no need to customise – this is a one-size-fits-all genetic prophylactic! That’s the theory, at least…

The statistical modelling, taken at face value, suggests that this approach could be hugely beneficial to individuals, and, by extension, to the health of the population as a whole (if widely deployed). The question is whether we should take it at face value. The modelling is based on lots of assumptions, which the authors, to their credit, take some pains to spell out. If these assumptions are not valid, they may undermine the predicted efficacy. Moreover, there are additional safety concerns that are just not part of the model at all – they arise due to possible pleiotropic and “epistatic” (non-additive) effects that are highly unpredictable.

Is a simple linear model appropriate?

The first assumption with this approach is that we know what the causal variants are at the relevant sites in the genome. It turns out that most of those SNPs are not doing anything themselves – they just act as tags for bits of chromosomes that tend to get co-inherited in chunks. The real causal variant is usually hanging around somewhere nearby, but it is often difficult – I mean really difficult – to identify the true culprit. So, even after GWAS has identified associated SNPs, there’s still a lot of work to be done to identify the causal variants, with no guarantee of success. But, for the sake of argument, let’s imagine a future world in which that work has been done and we have perfect knowledge of the causal variants at each site.

Even with that, a more crucial assumption in the modelling used is that the effects of these different genetic variants will simply add up, in a statistical sense. That is, that the decrease in risk due to any individual variant will be independent of the presence or absence of other variants in the genome. You can just keep adding protective variants to a genome and you’ll keep seeing the same relative reductions in risk. Is this a reasonable assumption? From one perspective, yes, and from another, not at all.

The difference hinges on whether we are talking about “statistical epistasis” or “biological epistasis” (reviewed here and here). Since people first started doing experimental genetics with model organisms – pea plants, fruit flies, bacteria, yeast, nematodes, mice – they’ve recognised that the effects of one mutation can be suppressed or enhanced by the presence of other mutations. They called mutations that suppress the effects of other mutations “epistatic” to them (“standing above”). This is just an extension of the better-known ideas of dominance and recessiveness, which capture the non-additive relationships of mutations at a single gene. Epistasis captures the same kind of non-additive effects of mutations in different genes.

This phenomenon is commonplace – ubiquitous, really. It is a mainstay of experimental genetics in model organisms, as geneticists use these non-additive relationships to work out functional relationships between different gene products, delineating biochemical pathways and regulatory networks. More generally, it is extremely common to find that the phenotype caused by some mutation is highly dependent on the “genetic background” in which it is found. For example, the phenotype may change, drastically in some cases, if a mutation is crossed into a different strain of flies or mice.

Overall, as I discuss at greater length here, the evidence from experimental genetics in model organisms suggests that the relationship between genetic variation and phenotypic variation is highly complex, non-linear, and unpredictable.

But now here’s a surprise. When people look for these kinds of non-additive effects in human genetics – at least in the genetics of common traits and disorders – they typically don’t find them. To be more precise, they don’t seem to make much of a statistical contribution at the population level of analysis. The same kinds of effects absolutely do hold when looking at rare mutations with large effects on disease risk. There are loads of examples of digenic interactions that result in disease or that modify the severity of symptoms. And loads of evidence that the effects of such mutations are sensitive to the genetic (polygenic) background (e.g., here, here, and here).

However, when we just examine the collective effects of (relatively) common genetic variants of the sort tagged by SNPs in GWAS, we typically see that a simple, linear model of purely additive interactions does a pretty good job of capturing most of the genetic effects across the population. That seems weird. If two or three mutations can show big non-additive effects, shouldn’t we expect thousands of them to show all kinds of higher-order effects, so that things might really get massively non-linear? Well, it seems not, for a few reasons.

First, the SNPs in question only have small effects by themselves (because mutations with big deleterious effects are rapidly selected against and never rise to an appreciable frequency in the population). This means they’re also less likely to have strongly deleterious epistatic effects – just because they’re not doing much, biologically speaking. So, the class of genetic variants that dominates the genetic architecture of complex traits in natural populations is very different from experimentally induced mutations in model organisms, or from rare (recently arising) mutations in humans that have very large effects by themselves (and in combinations).

Second, when there are rarer (but still relatively common) SNPs that do have non-additive pairwise interactions, the statistical contribution of these epistatic effects to the overall genetic variance across the population depends on both the size of these effects and the frequencies of the variants. The rarer they are, the less they’ll be seen in combination. So any specific epistatic effect may only show up in some very small number of individuals, and thus will not make a big contribution to how genetic variation manifests across the whole population. In addition, the statistical method that is used to divide up the variance just prioritises the additive contribution, even though lots of non-additive interactions are included in that term.

As Visscher and Wray write: “Because of the dependency of the variance components on allele frequency, a strong deviation from additive gene action can result in mostly additive genetic variation.”

All of these factors mean that simple, linear models of genetic “liability” for common disorders actually do a reasonable job of treating the collective effects of common genetic variation across the population. It is that tradition that underpins the modelling in the current paper. The problem with this approach is that the proposed interventions would not be done on the population – they’d be done on specific individuals. So, we have to ask, for such individuals: would the intervention be as effective as predicted, and would it be safe?

The answer to both of these questions is: we just don’t know. In fact, it may be unknowable, without doing the experiment. For all practical purposes, it may be impossible to predict either efficacy or safety with any certainty. We simply can’t capture all the possible complexities of the biology of (genuinely) complex disorders with simple linear models based on population-averaged data.

Will the effects be as large as predicted?

The variants that Visscher and colleagues propose to introduce are specifically relatively rare, “protective” ones. Their modeling depends on the idea that their protective effects will simply add up. But now we’re in the space where biological epistasis really matters, because we’re talking about combining rare alleles in ways that don’t normally occur in the population. The only estimate we have of their protective effects is averaged across genomic backgrounds that do not carry many of the other rare variants.

Based on experience from model organisms (see here for detailed examples), it seems very possible, even highly likely, that combining these kinds of rare variants would result in diminishing returns – that is, their protective effects would be reduced if other protective variants are already present.

As Sackton and Hartl wrote in a very relevant article entitled “Genotypic Context and Epistasis in Individuals and Populations” in 2016:

“Paradoxically, the effects of genotypic context in individuals and populations are distinct and sometimes contradictory. We argue that predicting genotype from phenotype for individuals based on population studies is difficult and, especially in human genetics, likely to result in underestimating the effects of genotypic context.”

Such effects would obviously reduce the efficacy of the intervention, as Visscher and colleagues concede: “Quantitatively, the effect of epistasis would be similar to what was modelled for G x E interactions (Fig. 2), where a reduction in actual outcomes compared to what is predicted in the absence of G x G interactions.” However, one could still argue that even reduced gains would still be worth pursuing, as long as there are no appreciable risks. The problem is that the authors do not fully explore the possible dangers of their proposed scheme.

Would it be safe?

As stated above, the first challenge will be to correctly identify the causal variants linked to our rare “protective” SNPs. This is not trivial, and experience has shown it is quite possible to get it wrong. Editing the wrong variant (one that is not in fact protective for the disorder, but which may have other effects) would obviously be a potential risk.

But, for the sake of the thought experiment, let’s presume we’ve been able to find the right ones. The causal variants will presumably also be rare, perhaps a good bit rarer in fact. The problem with introducing rare variants, as stated above, is that they may be rare for a reason. That is, a variant that is statistically protective against some common disease may also have negative (pleiotropic) effects on other phenotypes and may consequently be under negative selection. After all, for many of the linked causal variants, we would be talking about replacing a much more common, presumably “wild-type” (or ancestral) version of the DNA sequences with a more recently arising mutant version. They might have some protective effect but they might also influence all kinds of other things. Recent results in humans reinforce this expectation.

If all we are basing our models on are statistical data for the diseases themselves, then we will have no window onto these other possible effects – they will be unknown unknowns. But the problem gets even worse when we consider the possibility of epistasis and pleiotropy acting together. Not only do we not know what possible negative effects these “protective” variants might have on their own. We really don’t know what negative effects they might have in combination. And there’s no way we could find out without doing the experiment, because it’s quite likely that no one on earth carries all of the rare versions at both copies of all the ten sites being edited. So now we're talking about unknowable unknowns.

Visscher and colleagues acknowledge the potential for negative pleiotropic effects: “Pleiotropy is the norm in genetics. Variants associated with decreased disease can also be associated with other diseases and traits and may increase their risk.” But the way that they assess the possible consequences of such effects is frankly troubling. They consider protection against disease as increasing fitness and other possible negative effects as simply decreasing fitness. They then state that overall gains in fitness will be reduced if negative effects arise:

“We modelled a possible deleterious effect of [heritable polygenic editing] on fitness using a model of stabilizing selection (Supplementary Note 2), and the results are shown in Supplementary Fig. 3. These results imply that the fitness of edited genomes can be substantially reduced if the phenotypic change is large and the trait is under a strong stabilizing selection. The consequences of pleiotropy mean that the actual effects of HPE on disease prevalence will be less than those predicted in this model, and the overall effects on quantitative traits not as strong as predicted.” (My emphasis in bold).

This framing amounts to saying that there is some level of acceptable casualties (individuals with “reduced fitness”) from these interventions, if the overall population gain is still positive. These complicated and unpredictable negative effects are just seen as reducing the overall efficacy of the intervention across all the edited genomes. This seems very much like a move from concerns of individual welfare to concerns of public health, and even from public health to eugenics.

It sounds very like the way animal breeders might talk about the effects of selective breeding, where they can accept undesirable side effects for some individuals, in order to get a change in the average of some desired trait in the next generation. Actually, even some in the animal genetics community are not comfortable with the idea that we should engage in widespread genome editing of livestock animals, for the same reasons – we just don’t know enough about the genetic architecture of complex traits.

Visscher and colleagues do consider the issue of safety, among various ethical concerns: “Heritable polygenic editing (HPE) introduces new combinations of variants that can be dangerous or unsafe. It would be unethical to impose this uncertain risk on future generations”.

Their response is: “It is vital that any use of HPE be supported by rigorous safety data and have a clear justification through a risk/benefit balance. In addition, natural reproduction generates new combinations of variants. One strategy is to limit HPE to variant combinations already seen in existing populations.”

And they go on to conclude that: “Before such applications are considered, it is important to conduct further research on the effect of polygenic variants on individuals in natural populations, including the lifelong consequences of carrying rare protective alleles.”

That cautious attitude is certainly warranted, but is a far cry from the modeling they present and the sensational gains they claim are possible. If you don’t think editing all those sites is a good idea, why present that as the centerpiece of your paper?

Sackton and Hartl discuss the relevance of genetic complexities for genome editing specifically. They end their article with this statement: “Probably the best that one can do under the circumstances, using present methods, is to make predictions based on the main, additive effects of alleles, recognize the uncertainty of such predictions, and hope for the best.”

“Let’s edit your baby’s genome and hope for the best” is not exactly a compelling marketing slogan. It’s certainly not a responsible medical attitude, where the dictum “first do no harm” is supposed to hold.

Eugenics and public attitudes to genetics

Overall, Visscher and colleagues present a fantastical scenario, based on all kinds of technical assumptions that may never be realized – primarily that CRISPR editing can be made completely precise and that we will be able to identify causal variants linked to associated SNPs. They go on to model the effects of gene editing with a simple, linear model that we know will not capture the possible biological complexities in individuals. And – in some places in their paper, at least – they treat possible negative consequences (real harms to real individuals) as simply reducing the efficacy of their proposed interventions across the pool of edited genomes.

It’s really not clear to me what the aims of the authors are in publishing this kind of speculative analysis. It seems the height of arrogance and statistical hubris to assume that their simple linear models based on population-averaged data can capture the biological complexities of genotype-phenotype relations in individual human beings. (Though the underlying sentiment certainly aligns with the zeitgeist in the tech world these days).

While in some places Visscher and colleagues strike a cautious tone, in others they explicitly argue for the uptake of this kind of approach, in terms that are frankly eugenic:

“In the long term, there may be an obligation to pursue and develop technologies such as HPE. Mildly deleterious mutations that escape natural selection because of better medical care are predicted to accumulate in the gene pool (62). Previously published models suggest that the effect of this ‘genetic load’ might manifest itself as physical and mental deterioration in only a few generations (62). However, this concept is controversial, and the conclusions are debated (63–65). If we take seriously the idea of leaving future generations in a better state than the current generations, then we have reason to provide them with the preconditions for a good life. This includes access to clean water, unpolluted air, education and shelter, and may include the use of HPE to lower the genetic risk of disease.”

This concern with ‘genetic load’ is a mainstay of eugenic thinking, but there is no good evidence that it is actually increasing or that we are facing the inevitable degradation of the human genome if we don’t intervene somehow. Appealing to it here in this offhand way, without the nuanced discussion it requires, seems highly irresponsible, especially when eugenics tropes are reappearing with increasing regularity in the public sphere.

Genetics will continue to have a greater and greater impact in our societies. We will need to have all kinds of ethical discussions about what it means and, as scientists, do lots of work to explain the complexities and nuances to the wider public. Indulging in the kind of statistical fantasies presented in this paper is unlikely to help these efforts.

Note that the opinions expressed above are my own and should not be attributed to the co-authors of our commentary. We conclude that piece in more measured fashion:

“Meanwhile, other technologies in reproductive genetics are already available or are around the corner, including embryo, fetus and newborn whole-genome sequencing. Each comes with profound clinical, ethical and societal questions. Upcoming population-scale genome-sequencing efforts raise further urgent questions of privacy, stigmatization and discrimination. Is it wise to distract stakeholders, including the public, with a technology that is still a long way off at best, and might never actually be safe?”

The Justice Algorithm

noreply@blogger.com (Kevin Mitchell) — Wed, 30 Oct 2024 20:50:00 +0000

Why do we need judges? Why do we leave important decisions to flawed, biased, distracted, even corruptible human beings? Couldn’t A.I. do this job so much better now? So much more cleanly, precisely – without all that messy, human subjectivity? Couldn’t we just submit the evidence to a great big algorithm to determine guilt or innocence? Or, if that involves too much ambiguity, at least to determine appropriate sentencing, taking all appropriate factors into consideration? Shouldn’t there be a single right answer that can be reached in each case?

After all, we have, in most jurisdictions, a constitution that lays out our moral and societal values and guiding legal principles. Of course, these aren’t specific to any situation, so we also have a set of laws that dictate very clearly what’s allowed and what isn’t. Admittedly, we keep having to make new ones to keep up with a changing world, but they’re as comprehensive and up-to-date as our political systems allow. We know what kinds of things are crimes, and if someone’s done one, they’ve done one.

Now, it’s true that the laws can’t cover every particular instance that would constitute each type of crime, but that’s where we can lean on decades and decades of legal precedent. We’ve seen it all at this stage – we’ve been meting out justice for centuries, after all!

All we need is the constitution, our sets of laws, and the body of legal precedent to parameterise our algorithm. Then we can just input the facts of the case and the machine should churn out the correct verdict and sentence. I suppose there might be some disagreement on what constitutes a “fact”. But we could handle that by attaching a certainty parameter to each one. Shouldn’t be a problem.

Admittedly, the considerations for sentencing get a bit complex. There are obviously a lot of factors to take into account, when thinking about what sentence someone deserves. The good news, though, is that body of legal precedent is so extensive that it’s very unlikely we’ll encounter a new factor we haven’t seen before. Once we have the weights for each factor included in our algorithm, we should be all set!

I mean, maybe there would be some scenarios where the weight you put on one factor would depend on the context of some other factors. But that shouldn’t be a problem either. We’d just need to include some second-order weights to tweak the first-order ones based on those kinds of contextual factors. And, I guess, maybe some third-order ones. But, honestly, how many combinations could there be? With all of legal history at our disposal, we should be able to match all possible contingencies, right? Shouldn’t we? Our program might get a little unwieldy, admittedly, but it’s not like we don’t have the spare compute and energy capacity. And I guess it might take quite a while to churn through all the necessary computations, but, hey, who’s in a rush?

And if an algorithmic program – some set of computational steps that we run through in series – can’t handle the complexity, maybe we could turn to a machine learning model that we could just train on loads of past cases. We’d just need to define the right information to give it and then let it figure out the weights for itself. It’d just be a great big exercise in pattern recognition. Again, once beyond some level of training, the past should be a certain guide for all future cases.

The important thing is that this process should be deterministic – no messy subjectivity or variability! Just clean computation. Each set of facts (weighted by certainty) and relevant circumstances (weighted by context, in ways that are themselves weighed by context) will reliably produce a single outcome.

That obviously doesn’t mean that there will be some kind of infinite look-up table containing a verdict and sentence for every conceivable combination of facts and circumstances, of course – that would be absurd! It’s just that our conditioning on the past would be so comprehensive that the result of the computations for any new scenario would be effectively fixed in advance. You’d just have to run the algorithm to find out what it is. The answer would already sort of “exist”, in potentia, as the philosophers say. The Algorithm will have spoken. By preprogramming or pretraining our machine, we should be able to fit any new scenario and know exactly what the right answer is! Huzzah!

I’m hoping by now that most readers will have detected just the subtlest note of sarcasm in the preceding arguments. And some of you may have figured out that I’m using the specific case of the legal system as an allegory for cognition more generally. (Though we’ll return to the Justice Algorithm itself at the end).

What prompted this analogy is a recent debate with Robert Sapolsky, in which he argued (based on his book “Determined”) that we have no free will in any moment – that our behavior is always, well, determined. The idea is that our brains and minds are shaped by all kinds of prior causes, from millions of years of evolution, to our own genetics and brain development, to all the experiences that have happened to us over our lifetimes.

All of those factors have been shown individually to have an influence on our patterns of behavior. No one disputes that. But Sapolsky makes the much stronger claim (without evidence, frankly) that the collective effect of all those prior causes doesn’t just influence but completely determines our behavior. Not just our patterns of behavior – our actual behavior in every specific situation we ever find ourselves in across our whole lives.

The question is: where does that leave you? What is there for you to do in this process? Under this view, your brain just constitutes a great big preconfigured algorithmic structure that is going to churn through its inputs (based on the available information about the world outside and your current states and motivations) and spit out a single, definitive answer, no matter how novel the situation. The equations will always have a unique solution.

To spell out the analogy, evolution would be loosely equivalent to the constitution, and, together with your own genetics and brain development, could effectively give you some general predispositions and behavioural principles. These would, however, not be very specific to any given situation. As described in my book Innate, they would just be a set of basic tunings of decision-making parameters (like risk aversion or reward sensitivity or novelty salience – that kind of thing). Those tunings don’t have any “content” – they’re not about anything in particular. Your individual history would be like the more specific laws and the huge body of legal precedent – everything that you have learned from the past that can be used to guide your behaviour in the future.

The problems with this view of deterministic decision-making are, I hope, obvious from the analogy. We can’t just run any set of new information through a totally predefined cognitive algorithm, because the parameters of the algorithm are always massively context-dependent. And we can’t prestate all the relevant first- and second- and third-order weights because the space of possible combinations across all scenarios we might encounter is effectively infinite and unknowable in advance. This kind of combinatorial explosion makes the problem computationally intractable.

That’s not to say that a lot of our behaviour isn’t fairly automatic. In many familiar or simple scenarios, we know what to do based on a simple set of habits and heuristics. No thinking required. But once things get more novel and more complex, we can’t just submit the information to a preconfigured algorithm. We have to figure out how the algorithm should be configured. We have to work out how to weigh up all the various factors, on the fly, in real time. In other words, we have to do exactly what judges do – make a judgment!

Not just in real time, either – we have to do it in good time. And we don’t have infinite compute or unlimited energy – we just have our limited brain, selected for efficiency, not precision. Indeed, in many cases, it’s not like there actually is a right answer, waiting to be found. Given the limited information that an individual has at any moment, and all the potentially conflicting goals over which they are trying to optimise, simultaneously, there will often be a range of possible actions with indistinguishable predicted utility. This is the well-known principle of “bounded rationality”. We can’t always reach a definitive solution – sometimes we just have to pick one without knowing in advance if it will prove to have been the best choice. We’re not omniscient and we’re not clairvoyant.

To navigate through a world that is constantly throwing up novel scenarios, we need to do work to judge the relative salience of different factors, relative to our suite of current goals. We need to infer not just what is out in the world, but which elements matter to us – what we should care about and pay most attention to and weight most heavily, given our current state, our ongoing projects, whatever behavioural agendas we are pursuing, and so on. In short, we need to make decision-making computationally tractable. John Vervaeke and various colleagues have called this process “relevance realisation” and have presented some biologically plausible ways in which it can be achieved, allowing organisms to make their way pragmatically in an open-ended world.

This will inevitably involve lots of heuristics, rough estimates, and other means of satisficing – that is, trying to satisfy the myriad demands on behaviour, based on all kinds of competing considerations, in a reasonable amount of time with limited resources. We don’t even try to reach a right answer, because there isn’t one – we just have to reach one that is good enough.

There is thus no reason to think that all these cognitive processes will be deterministic, or even algorithmic, in the strict sense of that word (implying a predefined procedure following a set of discrete operations carried out in series, usually completely reproducibly). There are, in most situations, lots of things we could do and lots of ways we could do them. In fact, the only way that these cognitive-level processes could be deterministic is if the underlying neural and physical processes were completely deterministic. And we know that they’re not.

Brains are wet, messy, noisy places, full of tiny, jittery components. That inherent randomness of low-level goings on is both a challenge and an opportunity. It means, first of all, that the future evolution of the system is not fixed by the current microscopic state and the laws of physics – lots of things could happen. The challenge for the organism is to make happen what it wants to happen. But this low-level under-determination also provides an opportunity: to evolve control systems that can, by virtue of their macroscopic configuration, constrain how the system evolves.

In my book Free Agents, I lay out the case for this view and describe the evolutionary trajectory that led to ever-increasing agency along our own lineage. One important element of this framework is that the workings of the system are causally sensitive to the meanings of neural patterns, rather than the low-level details of their momentary instantiations.

If this view is on the right track – if the neural computations are causally sensitive to semantic content, rather than detailed syntax, and those semantics relate to organism-level concepts, and all that information is integrated in a hugely contextually interdependent way, and is used to direct behavior over nested timescales, in ways that cannot be either algorithmically or physically pre-specified, based on criteria configured into the circuits derived from learning, which embody reasons of the organism and not any of its parts, then I would say that *just is* the organism – as an integrated self with continuity through time – deciding what to do.

To me, this picture matches the phenomenology of making decisions under conflict much better than the idea that our brain just spits out an answer. We really have to think about things – it’s effortful, even exhausting sometimes. And that kind of deliberation is a process, extended in time – one that we actually have a window on. We can consider whether the reasons we’re operating with are good reasons, whether the information we have is trustworthy or consistent, even whether our motivations are the types of motivations we think we should have. (There’s much more on that final point in the book). The point is that these are not just processes happening within us – we ourselves, as selves, are very actively making judgments.

Now, I promised to return to the idea of the Justice Algorithm. I offered it as an argument in absurdum – it’s clearly such a simplistically reductive idea in the context of the legal system that it prompted the intuitions I wanted in relation to natural cognition. Welp, seems the joke’s on me! AI systems are already in use in many criminal justice systems, using individuals’ profiles on hundreds of metrics to make predictions on risks of flight, recidivism, violence, sexual offenses, and so on, which are then used to guide decisions on bail, sentencing, parole, admission to rehabilitation programs, and other important outcomes.

At present, those AI algorithms are not making any decisions by themselves – they’re just providing information that judges or other workers in the justice system can use to help guide decisions. There are, however, huge questions over whether that information is useful, and, more importantly, over whether its application is fair or actually codifies all kinds of biases. I’ll let the experts speak to those issues, in the articles listed below. In the meantime, I can only hope that we are not heading towards a future where “Computer says jail”.

A ‘black box’ AI system has been influencing criminal justice decisions for over two decades – it’s time to open it up

https://theconversation.com/a-black-box-ai-system-has-been-influencing-criminal-justice-decisions-for-over-two-decades-its-time-to-open-it-up-200594

AI is already being used in the legal system – we need to pay more attention to how we use it

https://theconversation.com/ai-is-already-being-used-in-the-legal-system-we-need-to-pay-more-attention-to-how-we-use-it-205441

AI is sending people to jail—and getting it wrong

https://www.technologyreview.com/2019/01/21/137783/algorithms-criminal-justice-ai/

Artificial intelligence (AI), data and criminal justice

https://www.fairtrials.org/campaigns/ai-algorithms-data/?gad_source=1&gclid=Cj0KCQjwsoe5BhDiARIsAOXVoUvuuEFhKSmGxzcM4eFAM-a34CxqOkxFiE674gM_guXNjHveRboBMBUaArwiEALw_wcB

Undetermined - a response to Robert Sapolsky. Part 4 - Loosening the treaties of fate

noreply@blogger.com (Kevin Mitchell) — Mon, 22 Jan 2024 19:25:00 +0000

In Part 3 of this series, I argued that organisms really do think about what to do, really do come to their reasons by reasoning, and really do make decisions, in ways that cannot be pre-determined.

If the neural computations are causally sensitive to semantic content, rather than detailed syntax, and those semantics relate to organism-level concepts, and all that information is integrated in a hugely contextually interdependent way, and is used to direct behavior over nested timescales, in ways that cannot be either algorithmically or physically pre-specified, based on criteria configured into the circuits derived from learning, which embody reasons of the organism and not any of its parts, then I would say that just is the organism – as an integrated self with continuity through time – deciding what to do.

I also argued that more fundamental principles of indeterminacy and emergence and organisation are the things that enable organisms themselves to come to be in charge of what happens

But what if the universe as a whole is deterministic?

How could mental states – or any states of the whole organism – have causal power over how the system evolves, when all the causes are supposedly located at the lowest levels? How can we escape from the confines of reductionism?

Sapolsky considers the question of physical pre-determinism at the lowest levels of reality and comes to a similar initial conclusion as I do, which is that the general consensus from physics is that the low-level laws of quantum mechanics are not deterministic:

For our purposes, the main points are that in the view of most of the savants, the subatomic universe works on a level that is fundamentally indeterministic on both an ontic and epistemic level. (page 213)

However, he argues that the law of large numbers means that any such random events at the level of individual molecules will be averaged out across the system and will thus not manifest at levels we care about. Oddly, one of his arguments for this is that the system in question is not deterministic at those higher levels:

People in this business view the brain not only as “noisy” in this sense but also as “warm” and “wet,” the messy sort of living environment that biases against quantum effects persisting. (page 221)

The argument thus seems to be that individual random events at low levels get washed out by higher-level randomness:

…quantum effects are washed away amid the decohering warm, wet noise of the brain as one scales up. (page 238)

It’s not clear where, in this picture, the higher-level randomness is supposed to come from, but actually, for the purposes of the discussion of free will, it’s sufficient that it exists. The important point is that physical pre-determinism does not hold, at any level. This means the evolution of the system through time should not be completely pre-determined, but open.

Sapolsky still maintains a hard reductionist stance, however, and while he engages with the concepts of non-linearity in complex systems, he fails to make a crucial connection. He writes that the reductionist tradition of only studying linear relationships in isolation:

…guaranteed the incorrect conclusion that the world is mostly about linear, additive predictability and nonlinear chaoticism was a weird anomaly that could mostly be ignored. Until it couldn’t be anymore, as it became clear that chaoticism lurked behind the most interesting complicated things. A cell, a brain, a person, a society, was more like the chaoticism of a cloud than the reductionism of a watch. (page 145)

However, based on the literature on chaotic systems, he concludes:

Even if chaoticism is unpredictable, it is still deterministic. (page 148)

Now, it is true that many of the classic examples of chaotic systems – or, more precisely, computer simulations of such systems – are deterministic. The evolution of such simulations can be extremely sensitive to slight changes in the values of different parameters at far decimal points, which can have unexpected, but consistent, influences over future states.

However, that does not mean that chaotic systems in the real world – outside of computer simulations – must also be deterministic. If there is some real indeterminacy in the physical parameters of a system (which we’ve established there must be), and it has the kind of complex, non-linear dynamics that make it chaotic, well then it will be genuinely unpredictable, not just in practice, but in principle. The parameters describing its future states just won’t have truth values in the present, beyond some decimal point. I argue in Free Agents that the universe as a whole is like that. It operates a kind of just-in-time reality, where the future is radically open.

How could indeterminacy help?

A common argument goes like this: if my behavior is caused by deterministic physical events, then I don’t have free will. But if it’s caused by random physical events, then I also don’t have free will. Sapolsky rightly raises this challenge:

Even if quantum effects bubbled up enough to make our macro world as indeterministic as our micro one is, this would not be a mechanism for free will worth wanting. That is, unless you figure out a way where we can supposedly harness the randomness of quantum indeterminacy to direct the consistencies of who we are. (page 230)

This almost gets us where we need to go, but misses what for me is possibly the most crucial point in this whole debate. For agency to emerge and be exercised, living organisms don’t have to harness the indeterminacy – they just have to take advantage of the causal opportunities it presents.

As my student, Henry Potter, puts it: “Indeterminacy is little more than a pre-condition for agency. It is the background on which a story of agency can be built, not a leading character in that story itself.”

However, Sapolsky continues:

I see two broad ways of thinking about how we might harness, co‑opt, and join forces with randomness for moral consistency. In a “filtering” model, randomness is generated indeterministically, the usual, but the agentic “you” installs a filter up top that allows only some of the randomness that has bubbled up to pass through and drive behavior. In contrast, in a “messing with” model, your agentic self reaches all the way down and messes with the quantum indeterminacy itself in a way that produces the behavior supposedly chosen. (page 231)

This is a nice distinction, but I don’t think either of these models is apt, and a third possibility is not considered. We do indeed use “filtering” mechanisms all the time – that’s how we evaluate possible actions and select one, for example. But this is not, except under certain circumstances where it’s useful, allowing some of the randomness “to drive behavior”. And I agree with Sapolsky that the “messing with” model makes no sense. We don’t need top-down causation to reach a ghostly tendril down and control what would otherwise be a random quantum event.

The important point is much more fundamental. I’m going to take the odd step of quoting him quoting me:

In the words of Irish neuroscientist Kevin Mitchell, “indeterminacy creates some elbow room. . . . What randomness does, it is posited, is to introduce some room, some causal slack in the system, for higher-order factors to exert a causal influence” (my emphasis). (page 235).

He seems to take me to be arguing here for a “messing with” sort of mechanism, but I’m doing quite the opposite. The causal influence that is exerted is on how the whole system evolves at macro levels. The organism doesn’t care how the system evolves on the micro level. Those details are not relevant.

As physicist and philosopher George Ellis puts it: "Much randomness occurs at the molecular level, which enables higher functions to select lower level outcomes according to higher level needs."

We don’t need (and generally don’t want) specific random events to determine the outcome. We just need some pervasive indeterminacy to exist such that the low-level forces under-determine the outcome, leaving some room for higher-order organisation to emerge, with causal efficacy over how the system evolves (because the system has been configured in functional ways by natural selection). This is not “messing with”, this is taking advantage of.

Sapolsky disagrees with philosophers Robert Kane and Christian List:

Robert Kane states the same: “We think we have to become originators at the micro- level [to explain free will] . . . and we realize, of course, that we cannot do that. But we do not have to. It is the wrong place to look. We do not have to micro-manage our individual neurons one by one.”

So these free- will believers accept that a neuron cannot defy the physical universe and have free will. But a bunch of them can; to quote List, “free will and its prerequisites are emergent, higher-level phenomena.” (page 192)

While I also disagree with the compatibilist version of these arguments (where fundamental indeterminacy is deemed to be unnecessary for high-level freedom), the general point is sound (especially if we can just help ourselves to fundamental indeterminacy, because that’s our best understanding of the relevant physics).

We really don’t have to micro-manage our individual neurons – that’s the whole point. We don’t have to care what every neuron is doing because the system is configured to be sensitive to the meaning of high-level patterns, not the precise details of every action potential or every synaptic vesicle fusing or ion channel opening.

Emergence and downward causation

The idea that an organism can have causal power, as a whole entity, as a self, depends on concepts of emergence and downward causality. It’s just not possible to grasp the main argument without engaging with these concepts. It is thus disappointing that Sapolsky mostly simply chooses not to:

Thus, a lot of people have linked emergence and free will; I will not consider most of them because, to be frank, I can’t understand what they’re suggesting, and to be franker, I don’t think the lack of comprehension is entirely my fault. (Pages 192-3)

Instead, he argues that:

Emergent systems can’t make the bricks that built them stop being brick-ish (page 202).

That latter point is true, but misses the larger idea, which is simply that the way a system is organised can (non-magically) constrain the behavior of its components, without changing their individual nature.

He states:

…the core belief among this style of emergent free-willers is that emergent states can in fact change how neurons work, and that free will depends on it. It is the assumption that emergent systems “have base elements that behave in novel ways when they operate as part of the higher-order system.” (Page 201).

But if “emergent states” means, for example, what you’re thinking about, then of course that alters how neurons work. That is precisely how nervous systems as a whole work – by one neuron altering how another one is working. And by population dynamics affecting how each individual neuron is working. And by patterns at one level – patterns that mean something – setting the context and criteria for how neurons at another level work. This isn’t magic.

In the same way, the electrons in your computer are constrained by the particular software that’s running. This doesn’t change the properties of the individual electrons, but it absolutely does mean that electrons are behaving differently when they’re in a microprocessor than when they’re not.

It’s thus a shame that Sapolsky does not engage with the kind of literature – by people like Alicia Juarrero and George Ellis, for example – that provides exactly the resources we need to understand how complex systems can be organised in functional ways, where constraints and history are just as crucial to understanding causation as local, instantaneous physical forces.

Mental causation

Perhaps the key stumbling block in understanding how things can be up to us, in a meaningful sense, is the idea of mental causation. Descartes’s dualism left us with the problem of how mental goings-on could possibly influence physical goings-on (discussed in Part 1). The whole point of positing the mental realm was to free our immaterial thoughts from physical determinism, but this does not explain how those immaterial thoughts can then push material stuff around.

The solution is to realise that thoughts are not immaterial. They are entailed by meaningful patterns of neural activity. Now, you might say that that proves Sapolsky’s point – that what happens in the brain is just determined by these physical patterns. But in fact the workings of the brain are not sensitive to the details of those patterns – they are sensitive to what the patterns mean.

The elements of cognition are not neural firings. They’re higher-order patterns of neural activation that represent beliefs and desires and goals and possible actions. Those patterns are “multiply realisable” – they are macrostates that can be realised by many different microstates. The brain is configured to be causally sensitive to these high-level macrostates – the details of the microstates, in contrast, are arbitrary and incidental.

So, we really do think. Cognition is not an epiphenomenon. Mental states are instantiated by some neural states but cannot be identified with or reduced to them. Neural states have causal influence on the system by virtue of what they mean. Conversely, having a thought, with some particular “content” has causal efficacy in the system because of this physical instantiation. No magic required. Just meaning.

Undetermined - a response to Robert Sapolsky. Part 3 - Where do intentions come from?

noreply@blogger.com (Kevin Mitchell) — Fri, 12 Jan 2024 13:55:00 +0000

In his book Determined, Robert Sapolsky argues that our intentions arise in a completely deterministic fashion from the combined effects of all the prior causes that have acted on us, right up to the moment of action. He contends (i) that our intentions determine what we do, and (ii) that we have no control over their formation – they just appear when we are confronted with each successive situation we encounter. Referring to a classic turn-back-the-clock kind of thought experiment, he says:

But no matter how fervent, even desperate, you are, you can’t successfully wish to have wished for a different intent. And you can’t meta your way out— you can’t successfully wish for the tools (say, more self- discipline) that will make you better at successfully wishing what you wish for. None of us can. (page 46, original emphasis)

Here, Sapolsky seems to be arguing for psychological determinism. Your behavior at any moment is fully determined by the sets of reasons that you bring to the situation. So, echoing Arthur Schopenhauer, you can (indeed you must!) “do what you want”, but you can’t “want what you want”.

In a chapter entitled “Where Does Intent Come From?”, Sapolsky catalogues all the prior causes flowing seamlessly into each other:

So on and so on. Each moment flowing from all that came before. And whether it’s the smell of a room, what happened to you when you were a fetus, or what was up with your ancestors in the year 1500, all are things that you couldn’t control. A seamless stream of influences that, as said at the beginning, precludes being able to shoehorn in this thing called free will that is supposedly in the brain but not of it. (Page 80)

Note first, as discussed in Part 2 of this series, that the studies supposedly demonstrating these influences are not, in my view, reliable (at all) and give a misleading impression of the strength of these effects. And note also the dualist framing, pointed out in Part 1 of this series, requiring free will to be “in the brain but not of it”. He continues in a similar vein:

In order to prove there’s free will, you have to show that some behavior just happened out of thin air in the sense of considering all these biological precursors. (page 83)

For Sapolsky, you are thus simply the product of all the passive influences that have moulded you to be the person you are. The way you are currently constituted then fully determines your actions in every possible scenario you might encounter, without you, as a whole self, really being involved in that process – specifically, none of what happens is up to you. You are, in essence, pushed around by your own reasons, none of which you had any hand in choosing.

As we will see, this kind of psychological determinism reduces to the claim that all these prior influences have collectively caused your physical brain to be configured in the way that it is, which then physically necessitates subsequent states, resulting in this conclusion:

It is why it is anything but an absurdly high bar or straw man to say that free will can exist only if neurons’ actions are completely uninfluenced by all the uncontrollable factors that came before. It’s the only requirement there can be, because all that came before, with its varying flavors of uncontrollable luck, is what came to constitute you. This is how you became you. (page 84)

Sapolsky takes issue with compatibilist arguments that claim that even if you have no choice or control over what you do right now – because of the way your brain is configured, which embodies all kinds of policies and heuristics that will inform your current intentions – nevertheless you can be held responsible for having those policies and heuristics. Here, I agree with him – it seems incoherent of compatibilists to admit that at any given moment we have no real choice (because determinism holds), but then claim that in the past we could have made choices of what policies to adopt, which we can now be held responsible for.

However, Sapolsky’s own view – diametrically opposed to the compatibilist position – is entirely circular. The logic is that, because you never had control, you can never have control:

all we are is the history of our biology, over which we had no control, and of its interaction with environments, over which we also had no control, creating who we are in the moment. (page 85)

He argues that because choice doesn’t exist in any instant, then you – as a person, a self – could have had no control over your own dispositions. The argument thus rests on the thing it’s trying to prove.

If, instead, choice does exist in any moment, and you do have causal power over your own behavior, including the adoption of policies and heuristics that will inform future actions, then this view of you being an automaton passively shaped by forces outside your control is undermined. In its place we get a picture – and a naturalised scientific framework – of organisms steering their own course through the world, as best they can, deciding what to do in any moment, based on what they have learned from past experiences, in the context of their ongoing plans and commitments and agendas. In fact, it is precisely this combination of historicity and future-directedness that constitutes being a self, with continuity through time.

Where do our intentions come from?

Our intentions do not, in fact, just follow automatically from how we are currently configured and then control what we do, like commands to be executed in a computer. They are the outcome of our decision-making processes. That’s the point of decision-making – to figure out what we should do. Once we achieve that, it becomes our intention. This could be an immediate motor action, like moving our hand to switch on a light. Or it could be a longer-term project, like intending to finish university – a goal we adopt, which then provides context for our future and ongoing activities, which in turn provide context for our specific actions.

We don’t just go around reacting to stimuli in isolated instants. We manage our behavior proactively through time. We pursue agendas, we adopt policies and heuristics, we make plans and commitments, and engage in projects with long-term goals. Choosing an action is really choosing an objective – a desired future state of the world (including the state of our selves). We make the future what we want it to be – that is the point of action. But not just the immediate next time-point – we work to make things in the far future how we’d like them to be. This longer-term view of the management of behavior through time gives a better perspective on where our goals and intentions come from.

Achieving our goals requires sustained action – we usually have to carry out some series of activities over some period of time. If I have the goal of eating, I may have to spend time cooking a meal. For this activity to achieve my goal, I have to keep doing it. So choosing the goal to cook dinner, based on the motivation of being hungry, constrains and informs what I will intend to do over the course of whatever period of time it takes to do that task. I may intend to chop up an onion, and sauté some chicken, and make a lovely little soy and ginger sauce, and so on. Those intentions thus come from me having adopted this goal and from me thinking about the steps I need to take to achieve it.

Similarly, if I decided to play a round of golf, I would have, in the process, decided to intend to put the little ball in the hole. That intention came from me. When we decide to do something, a whole bunch of intentions have to come along in order to achieve that. So it’s not the case that while we can do what we want, we can’t want what we want. Our reasons don’t just reflect the pre-configuration of our brains in isolated instants. We come to those reasons by reasoning.

Those processes of decision-making must take into account our current state, the state of the world, the states we would ideally like both our selves and the world to be in, and then (i) suggest, and (ii) evaluate, options for action to find the one most likely to further the likelihood of those desired states. Now, you might argue that we don’t have control over the middle bit – the states we would ideally like our selves to be in. And, at a general and low level, that’s true – we’re wired to want to be well fed and watered, and warm and safe and loved and even entertained and fulfilled, and so on. Most fundamentally, we’re wired to want to survive.

But those kinds of basic evolutionary drives are not sufficient to determine our behavior. They are too general and context-independent. The same is true for the broad tunings of decision-making parameters that are reflected in what we call personality traits. We can’t rely on these general drives and tunings to tell us the best thing to do in any given situation. We have to be able to assess the particular situation we’re in and set more specific, context-appropriate goals, and select actions to achieve them. This entails cognition.

Why we need cognition

Cognition, roughly speaking, means using information to solve problems – usually to manage behavior over changeable environments. In some cases, a lot of the work of cognition is pre-done by evolution. For example, the biochemical configuration of a bacterium may effectively embody adaptive control policies, such as the tendency to move in certain directions, based on information acquired by receptor proteins about various substance that are out in the world. These kinds of systems can be reasonably complex – able to integrate multiple signals at once and mount a response that depends on many other contextual factors, including the recent history of the individual bacterium itself.

Even simple behaviors like chemotaxis thus entail holistic and integrative system-level control, aimed at enacting the optimal response. But what is the optimal response? In the bacterium, there is a manageable number of variables at play. Even if the context-dependent interactions are complicated, they’re not unworkably complicated. Evolution can pre-code a system of contingencies and context-dependent relations that are sufficient to deal with most scenarios that the organism will encounter, based on the limited range of scenarios its ancestors encountered.

That strategy has limits, however. It certainly won’t do for us. Humanity’s superpower – our whole evolutionary schtick – is cognitive flexibility. We don’t have hard-coded responses to every situation – we couldn’t have. There isn’t enough information in our genomes or our brains to pre-code defined responses to every kind of situation we may encounter. There are too many variables at play, too many relations to consider, too many goals to manage.

The solution is to decouple inputs from immediate action and instead gather information about what is out in the world and submit it to a central system, where it can all be considered, in the light of our stored knowledge about the world, and our current and ongoing goals, in order to try to decide on the best course of action to globally optimise over a host of variables all at once. In other words, we need to think.

In some cases, there may be one clear optimal course of action. But in many cases, there may be several, or none. The organism has to take in all kinds of sensory data, do its best to infer the causes of those data (i.e., infer what is out in the world), link that to its imperfect and incomplete knowledge, in order to update its model of the world with varying degrees of certainty attached to different elements, predict the outcomes and utility of possible actions, while the world and other agents are changing too, all while trying to satisfy multiple goals at once. That just is the organism deciding what to do, in real-time.

There seem to be two ways to take Sapolsky’s counter-argument, in relation to these processes of decision-making. Either, as put by my student Henry Potter: (i) “all of the system’s prior influences have hammered into it a neural and biochemical configuration that essentially embodies a look-up table for which action to perform under different circumstances. There is therefore no decision-making process going on. In effect, the decision is already made, it just needs to be ‘realised’ or ‘enacted’.” Or: (ii) processes of decision-making do have to happen in each new scenario, but they always have only one possible outcome – that is, they proceed algorithmically and deterministically.

The first interpretation seems impossible and I presume is not actually what Sapolsky has in mind. There is no way to pre-state all the scenarios an organism might find itself in and all the appropriate actions. But the second interpretation is no better – or, at least, it’s wildly speculative.

Resting psychological determinism on neural and physical determinism

Sapolsky seems to be arguing, without evidence, that all of those activities of reasoning and deciding are algorithmically deterministic – that there is only ever one possible outcome. This is basically a behaviorist and mechanistic position. It sees our brains as just one big (though very complicated) stimulus-response machine. Fundamentally, it denies that cognition or mental causation is real – that us thinking about what to do could have any causal influence in this physical system. It says the “contents” of our mental states don’t really matter – it is only the vehicles of those states that have causal efficacy in the system.

It takes these cognitive operations as really no more than the firings of certain neural circuits, which were in fact inevitable, at every moment in this sequence of events. This rests on the idea that the current configuration of our nervous system (which is the product of all those prior causes), plus the current incoming stimuli, deterministically result in a single next state of our nervous system, entailing some specific action. There is no choice. In every situation, all these antecedent causes will necessitate just one subsequent outcome.

There are two ways in which these processes could be algorithmically deterministic. The most obvious one is if the substrates they are instantiated in are physically deterministic. Then the whole argument just reduces to physical determinism, and the psychological framing of reasons and intents becomes an epiphenomenon. This is indeed the argument that many free-will skeptics make.

However, as we will see in Part 4, Sapolsky himself accepts that isn’t the case. He discusses the evidence that there is both fundamental indeterminacy in physical systems and pervasive noisiness in neural systems. So we shouldn’t expect any neurally instantiated process to be fully deterministic in its physical details.

The determinist’s move to escape from this challenge is, ironically, a determinedly anti-reductive one. It relies on the idea that all that noisiness of neural components and their physical micro-constituents is coarse-grained over – filtered or averaged out over the large numbers of components of the system. Then you could get an emergently deterministic outcome at the macro-level, with the noise simply evaporating. This would of course depend on the structure of the system – on the way that it is organised.

This move is doubly ironic, because, as I will explore in Part 4, it is precisely these factors of indeterminacy and emergence and organisation that enable organisms themselves to come to be in charge of what happens.

Part 1: A tale of two neuroscientists

Part 2: Assessing the scientific evidence

Undetermined - a response to Robert Sapolsky. Part 2 - assessing the scientific evidence

noreply@blogger.com (Kevin Mitchell) — Thu, 28 Dec 2023 11:26:00 +0000

In Part 1 of this series, I explored the different philosophical premises that Robert Sapolsky and I bring to the question of free will, in our respective books, Determined and Free Agents. Here, I will examine the scientific evidence that Sapolsky marshals to make his argument that all our decisions are fully determined.

Part 2

No, but yeah, but no – assessing the scientific evidence

Sapolsky presents an array of experimental evidence from studies of various kinds to support his claim that we are completely driven by all the causal factors in our past or intervening on us in the present. The word study appears 163 times in the text, in fact, and it felt a bit like being pummeled into submission at times. I’m all for providing experimental evidence to support one’s claims, but in this case, much of the supposed evidence is completely unreliable. The fields that are cited the most include social psychology, especially social “priming” experiments, candidate gene association studies, and certain kinds of neuroimaging studies. All of these suffer from very well documented methodological problems, known collectively as “questionable research practices”.

The problems stem from small samples, poorly defined hypotheses, excess researcher degrees of freedom (lots of ways to analyse the data), post hoc analyses, covariate mining (slicing the data in lots of ways to look for something significant somewhere), p-hacking, failure to correct for multiple tests, lack of independent replication, and the huge issue of publication bias. Collectively, these practices are empirically demonstrated to produce literatures composed almost entirely of spurious findings – apparently significant associations that really are just statistical blips. Because we only hear about the “positive” results, these bodies of literature can collectively give the impression of robustness and consistency, when in fact this results from biased reporting.

Here are a few examples from the text:

In one highly cited study, subjects rated their opinions about various sociopolitical topics (e.g., “On a scale of 1 to 10, how much do you agree with this statement?”). And if subjects were sitting in a room with a disgusting smell (versus a neutral one), the average level of warmth both conservatives and liberals reported for gay men decreased. Sure, you think— you’d feel less warmth for anyone if you’re gagging. However, the effect was specific to gay men, with no change in warmth toward lesbians, the elderly, or African Americans. (Page 47)

The study is presented as solid and reliable (with the implication that it also has some relevance for real-world behavior). But you should ask if the supposedly interesting “specificity” of the result was actually an unhypothesised and spurious finding from slicing the data multiple ways, and you should also ask if it has ever been replicated. Moreover, you can be sure that no one has read the results of other similar studies that may have been done that did not produce any such “positive” finding, because they never get published.

The whole field of social priming has been called into question as literally the poster child for the replication crisis. Even Daniel Kahneman, who was a huge proponent of this kind of work, has more recently had to concede that it is not robust and has called on the field to face up to this fact.

Here’s another example:

What best predicted whether a judge granted someone parole versus more jail time? How long it had been since they had eaten a meal. Appear before the judge soon after she’s had a meal, and there was roughly 65 percent chance of parole; appear a few hours after a meal, and there was close to a 0 percent chance. (Page 106).

This particular study is trotted out frequently, again to show how much our behavior – even on really important decisions such as this, made by people literally selected to be professional decision-makers – is driven by subconscious biological states, such as how hungry we are. This very famous example appears to actually be driven by the fact that defendants without representation, who are less likely to get parole, tend to have their cases scheduled right before lunch (because they’re quick).

The same unfortunate trend is evident across many other types of study cited. These include ones where people’s behavior is supposedly changed by exposure to various neural signaling molecules. For example:

Boost your oxytocin levels experimentally, and you’re more likely to be charitable and trusting in a competitive game. (page 54).

Notice first how this is presented not as a particular finding from a particular study, but simply as a fact with general predictive validity. There is a veritable cottage industry of studies involving spraying oxytocin up people’s noses and testing them on all kinds of behaviors. Again, the methods and the findings have been shown to be unreliable.

On the role of genetics in our behavior:

Consider the neurotransmitter serotonin— differing profiles of serotonin signaling among people [KM: due to different genetics] help explain individual differences related to mood, levels of arousal, tendency toward compulsive behavior, ruminative thoughts, and reactive aggression. (Page 72).

This section refers to results from so-called candidate gene association studies. Again, this methodology is no longer used because it is now well known to produce spurious findings, which have failed to replicate in much larger, much more systematic studies. The gene encoding the serotonin transporter, referenced here, is probably the most notorious example of this untrustworthy literature.

Finally, many neuroimaging studies are appealed to, which can seem to bolster the findings from social psychology experiments by referring to different parts of the brain “lighting up” differently under different conditions across different groups. The tortured phrasing of that sentence is intended to highlight the large number of variables and possible interactions in such studies that researchers can mine for significant findings. Here’s one illustrative example, with a typically convoluted and arbitrary set of interactions:

In another neuroimaging study, performance on a frontal task declined in subjects primed with pictures of spiders (versus birds); among African American subjects, the more of a history of discrimination, the more spiders activated the vmPFC and the more performance declined.

Again, recent work has highlighted how unreliable these kinds of studies are. In particular, studies that look across the whole brain for any kinds of differences between test groups are typically statistically “under-powered” by two to three orders of magnitude. That is, many of them have samples in the tens, when, to be reliable, they would need samples in the thousands or tens of thousands.

So, overall, the evidence base that Sapolsky draws on to emphasise the potency of all these supposed determinants of our behavior is, to put it bluntly, completely untrustworthy. This isn’t just my opinion. This is the opinion of researchers in these fields themselves, many of which, like human genetics, have completely overhauled the way they work to overcome the limitations of these kinds of “artisanal” studies.

But yeah…

That was the “No” part. But there is a “but yeah” to follow. Even though the types of studies that Sapolsky relies on are unreliable and unconvincing, this doesn’t mean such effects don’t exist. These factors are probably not as individually potent as he suggests, but it is still absolutely true that all kinds of prior causes influence our behavior.

For example, the fact that we can’t identify strong associations between individual genes and individual personality traits does not undermine the very strong evidence that such traits are partly heritable. (That is, that a sizeable proportion of the variance observed in such traits across the population can be attributed to genetic variation generally). It just means the relationship between genotypes and phenotypes is highly complex.

The same is true for neuroimaging experiments – the failure to associate complex psychological traits with variation in structure or function of isolated brain areas does not undermine the idea that such traits are associated with some differences in the brain. Again, it’s just highly complex.

And while most of the social priming literature is clearly artifactual, that doesn’t mean that we are not affected in the real world by all kinds of factors in the environment, some of which we may not be aware of.

All of these factors constrain (or inform) our behavior. However, if the existence of such factors is supposed to be a challenge to the idea of free will, it is only a challenge to an absolutist form. Some argue that having free will requires being absolutely free from the influence of any prior cause whatsoever. A quick scratch beneath the surface reveals how incoherent this notion is. A being that was free from all prior causes would have no reason to do anything, no evolved nature to begin with, no responsiveness to the environment, and neither future-directed goals nor memories of past experience to guide its behavior. It would just be a random-behavior-generator.

It would not in fact be a self with any continuity through time – and continuity through time is the very thing that defines selves. For any living organism, being a self entails doing work to constrain its component parts and processes to remain organised in the particular way that defines that individual. This is just as true at a psychological level as it is at a physical level. It is precisely our individual personalities and character traits, our memories, our ongoing projects and commitments, our habits and attitudes and policies that collectively make us who we are. Indeed, when people start acting “out of character”, it is often a sign of neurological or psychiatric illness.

So, we don’t have absolute freedom – we wouldn’t be ourselves if we did. What we have are degrees of freedom – some set of options available to us, informed by all these past and current factors. These influences only become a threat to a more reasonable conception of free will if they are taken to be complete determinants of what we do at any moment – if they reduce our degrees of freedom to zero. This is what Sapolsky asserts, but without evidence.

But no…

In discussing these various influences, stretching backwards over different timescales, Sapolsky is careful to state that each one of them is only an influence and not an absolute determinant by itself. That is good, because if any one of them (say your genetics) were determinative, it would rule out any role for any of the others (like your experiences).

But he does assert, completely speculatively and without evidence, that all of them collectively do completely determine our behavior, precluding the possibility of the organism itself playing a role in settling what happens. He claims that while we don’t yet have all the details, “we already know enough” to say there is not the smallest gap left in which free will could reside. (Note that this also means no further biological causes can be admitted, since we’re full up of causes already).

As an aside, but perhaps an important one, the use of the words “causes” in these kinds of discussions is highly loaded and likely to lead our thinking astray. It suggests a chain of necessitating events, deterministically driving behavior. And it seems to require the mythical beast – an “uncaused cause” – to arise somewhere to break this inexorable chain. But if you replace “causes” with “influences”, you get a much more open, flexible, ecumenical picture that much better captures the real nature of what’s going on.

Sapolsky doesn’t do that, however. His view is that all these influences collectively determine what we do at every moment. Clearly, this goes completely against the phenomenology of our everyday experience. It seems, most of the time, that we are in fact deciding what to do, that we do have options and we do choose between them, even if that is within a constrained set of possibilities. To assert otherwise is thus truly an extraordinary claim, which should be accompanied by extraordinary evidence. In the absence of any such evidence, Sapolsky relies on a series of arguments to try and make his case. One of those arguments – the central one, in fact – is that while we can do what we intend, our intentions are not up to us.

In Part 3, I will examine why that argument misses its mark.

Undetermined - a response to Robert Sapolsky. Part 1 - a tale of two neuroscientists

noreply@blogger.com (Kevin Mitchell) — Wed, 20 Dec 2023 18:57:00 +0000

Free will is in the air. Among neuroscientists at least, the question of whether we are in control of our actions has been attracting renewed attention of late, driven in large part by the successes of the field in laying bare the neural machinery of behaviour. It’s thus not a total coincidence that two books by neuroscientists on this topic – Determined, by Robert Sapolsky, and Free Agents, by yours truly – have been published so close together (both in October, 2023). What may be more surprising to readers is that we reach such divergent conclusions on the topic.

With thanks to Bill Sullivan

That we can survey the same evidence and interpret it so differently may suggest to some that the question of free will is not really an empirical one at all. Of course many of the issues are metaphysical, but both Prof. Sapolsky (Sapolsky hereafter) and myself think that the science of decision-making is at least relevant to these philosophical discussions. In short, he thinks that science definitively rules out the possibility of free will. I think it doesn’t. In fact, I think we can construct a perfectly unproblematic scientific framework that naturalises the concept.

A number of reviews have compared our two books (including The Times Literary Supplement, Nature, The Wall Street Journal, Undark magazine, 3 Quarks Daily, Nautilus, and others), and we also had the chance recently to debate the issue. But there were many points left unexplored and here I want to respond more directly and more thoroughly to some of the claims that Sapolsky makes in his book.

Where we agree

First, I should point out numerous areas where Sapolsky and I agree. A prominent view on free will among philosophers and those scientists who trouble themselves with the issue is compatibilism. This is the view that determinism holds (more on what this is taken to mean below), but that free will can be taken to be compatible with it. Or more precisely, that the concept of moral responsibility can be taken to be compatible with determinism. As espoused most forcefully by people like philosopher Daniel Dennett, this view accepts that there is only one possible future, but still argues that agents can be held responsible for their actions if they are acting for their reasons, even if they actually never have any choice in what occurs. Both Sapolsky and I find this view incoherent.

I also think it’s moot, given the evidence from physics that the universe is not in fact deterministic in that fashion. Sapolsky appears to agree with this and accepts that there is fundamental indeterminacy at the lowest levels. What he takes this to mean for determinacy at higher levels is a little less clear and is something I will come back to.

We also agree that there are all kinds of factors that influence our behaviour. My previous book, Innate, was all about how the wiring of our brains shapes who we are. We are definitively not born as blank slates, but have our own individual natures – innate predispositions that derive from our genetics and the way our brain happened to develop. Both Sapolsky and I agree that these predispositions affect the way our patterns of behavior emerge throughout our lives.

However, I take these effects to be distal, indirect, and non-exhaustive, which is why I used the word “shapes” rather than “determines”. As we will see below, one of the main points of disagreement is that I think of these and other factors as influences on our behaviour while Sapolsky sees them as absolute determinants of it.

Finally, I am very sympathetic with Sapolsky’s concern for what philosophers call “moral luck”. This is the idea that many of the things that people do in life – many of the ways in which their lives turn out – are due to factors over which they had no control and for which they deserve neither credit nor blame. This includes their natural endowments – talents, cognitive capacities, personality traits, and so on – as well as their social circumstances. Where we differ on this point is that I think it is perfectly possible to take such factors into account in moral or legal considerations, without having to take an absolutist metaphysical position that these prior causes preclude any capacity of decision-making whatsoever on the part of the individual.

Clearly, then, we see much of the evidence regarding the nature of the world and the nature of human beings as entities the same way. Nevertheless, we end up with different conclusions. One explanation for this could be that we start with different premises, perhaps implicit or tacit ones. It is thus worth exploring the assumptions and stances we bring to bear that inform our thinking.

Closet dualism

Sapolsky rightly criticises the position of dualism – the idea, inherited from Descartes, that there are two separate kinds of stuff, physical and mental. Under this view, the mental realm is where we make decisions, thus freeing us from the confines of physical determinism. This implies the existence of a “ghost in the machine” – an immaterial self that somehow inhabits your brain and body and pushes things around. Sapolsky derides this view as unscientific, which it obviously is, and rightly says that it provides no solution to the problem of free will – at least none that any scientist committed to the idea of physicalism (basically, no magic allowed) should accept.

However, his own position is dualist through and through. Right from the get-go, he frames the idea that a human being – as a whole entity – could be a cause of something as ludicrous and supernatural. On page 2, he writes, in relation to any behavior, that we can ask:

Why did that behavior occur? If you believe that turtles can float in the air, the answer is that it just happened, that there was no cause besides that person having simply decided to create that behavior. Science has recently provided a much more accurate answer, and when I say “recently,” I mean in the last few centuries. The answer is that the behavior happened because something that preceded it caused it to happen.

I guess I’d better explain the reference to turtles! This refers to an anecdote about a woman arguing with William James that he had his cosmology all wrong and the earth really rested on the back of a turtle. When he asked her what the turtle rested on, she said “another turtle”. And when he asked the inevitable question of what that turtle rested on, she replied: “It’s no use, Professor James. It’s turtles all the way down!” I gather Sapolsky is sympathising with the woman in the story, in the sense that every cause that we might give for an action inevitably has the same kind of regress of prior causes that it “rests on”. (There is also a hint here of the reductionist instincts that permeate the book – looking down instead of up).

The framing of the issues in the cited paragraph is firmly, though implicitly, dualist. First, there is the odd idea that a behavior occurring because the person “simply” decided to do something is basically identical to it having “just happened”, as if for no reason. But the whole point is that we can do things for our reasons, not at random. And as we will see, there is nothing “simple” about our decision-making processes. Nor is there any reason, a priori, to rule out the possibility of the person themselves being a cause of things – that’s the very question under debate.

Then there’s the implication that any such causation would require a miracle (the floating turtle) and is thus patently absurd. This only follows from the dualist position that when we say something is up to you, the “you” must be some kind of immaterial soul or ghost in the machine – as opposed to just your holistic self. Again, this rules out from the get-go a more naturalistic explanation – the very thing we’re after.

And finally, there is the assertion that science can replace – and indeed has replaced – such superstitious absurdities with the real explanation, which is couched in a linear chain of inexorable and inevitable causes of the respectable sciencey kind.

These themes are repeated throughout the book. On multiple occasions, Sapolsky implies that if biological factors are at work, these can’t provide the explanation or the vehicle for any real kind of free will, just by definition.

Each prior influence flows without a break from the effects of the influences before. As such, there’s no point in the sequence where you can insert a freedom of will that will be in that biological world but not of it. (Page 46)

The image of “inserting” a freedom of the will into the biological world, so that it is “in it, but not of it” is a clear articulation of a dualist position.

Again, on page 83:

In order to prove there’s free will, you have to show that some behavior just happened out of thin air in the sense of considering all these biological precursors.

Why “out of thin air”, as if this requires magic? This just assumes that the individual cannot be a cause of their own behavior – exactly the question at hand. If you set up the problem such that the only solution that would meet the threshold of free will is one involving this kind of supernatural interference in the normal run of physical causation, then of course you will never be satisfied by any naturalistic claim that explicates behavioural control as an evolved capacity of living organisms.

God of the gaps

This dualistic framing is also evident in Sapolsky’s repeated assertion that the only place where “you” can reside is in the gaps in our current knowledge of biology. Science is seen as revealing the real causes of what happens, at deeper and deeper mechanistic levels. As this occurs, there seems less and less for you to do, less and less room for any kind of holistic or top-down causation. All the causal power is located at the lowest levels (whatever they are taken to be) and the notion of higher-order causation by the self, as a whole entity, is eliminated in the process.

This is, of course, how atheist scholars characterise religion, as having to stage strategic retreats with every advance of science, taking refuge in whatever mysteries remain – hence, the god of the gaps. The implication is that science will eventually close all these gaps and the need to invoke a deity to explain anything at all will disappear.

Sapolsky’s eliminative reductionism similarly seeks a full causal explanation of behavior in the biological mechanisms that science is laying bare. This is thus the flip side of his implicit dualism – if the only thing that would be real free will is control by an immaterial self, apparently without involving any mechanism, then of course the more mechanisms biology reveals, the less room there is for the self to wield any influence, or even to exist at all.

My own position is one of non-reductive materialism. We don’t need to invoke any supernatural forces to explain how organisms can control their actions and how agents can have causal power in the world. There are plenty of resources available to us in the various sciences of systems that give perfectly naturalised concepts of holistic or top-down causation. Sapolsky considers some of these ideas in later chapters, but in a disappointingly dismissive way that, in my view, fails to fully engage with them. As such, I think he misses the very ideas we need to understand how causal agency can exist. It may be turtles all the way down, but it’s not turtles all the way up.

“Don’t look up!”

In eliminating the possibility of the whole self being a causal entity, Sapolsky offers a different challenge to proponents of free will, one that looks down, instead of up. In the first chapter, for example, he describes a scenario where a man pulls the trigger of a gun. He describes the set of causes of this event as the series of nerve impulses, working backwards from the ones that caused his trigger finger to contract, to the ones that caused those neurons to fire, and so on. (Note the purely mechanistic, driving framing – more on that later). He then says:

Here’s the challenge to a free willer: Find me the neuron that started this process in this man’s brain, the neuron that had an action potential for no reason, where no neuron spoke to it just before. (page 14)

The idea is that if you could show him “a neuron being a causeless cause in this sense”, then you would have proven free will. This just strikes me as such an odd set-up. Why would some neuron firing at random and thereby causing a behavior constitute an instance of free will? Why look for such a reductive explanation, at the level of specific neurons? We know that’s not how the brain works. He goes on to say:

The prominent compatibilist philosopher Alfred Mele of Florida State University emphatically feels that requiring something like that of free will is setting the bar “absurdly high.” (page 15)

I disagree. It’s not that the bar is too high. It’s just the wrong bar. We’re on the wrong playground. What he’s asking for would not in fact satisfy anyone’s concept of free will, including his own (because it would not give any causal power to the individual). He’s looking down when he should be looking up (but not up so far that the explanation has to sit somehow outside our heads in a ghostly self).

So, we differ in some of our philosophical stances in ways that provide at least some explanation for why our interpretation of the scientific evidence also differs so profoundly.

As to that evidence itself, most reviewers have taken Sapolsky’s presentation of it as a fair assessment of the profound influence that a whole range of biological factors can have on our behavior. Here, I am in an odd position: I agree that such factors can influence our behavior – of course they can – but I find the actual evidence that Sapolsky marshals to support this notion to be highly flawed. Indeed, the types of studies that he leans on to make this argument are literally ones that I (and many others) use as prime examples in lectures about irreproducible science and the “replication crisis”.

What questions should a real theory of consciousness encompass?

noreply@blogger.com (Kevin Mitchell) — Mon, 18 Sep 2023 11:54:00 +0000

Well, now! The consciousness field is all atwitter! A letter has been published, with 124 signatories, claiming that one prominent “theory of consciousness” – the Integrated Information Theory proposed and developed by Giulio Tononi, Christof Koch and colleagues over several years – is “pseudoscience”. That’s a serious charge to level in print, and one that I presume the authors of the letter did not make lightly.

The letter was a response to some of the media coverage around the COGITATE study – an adversarial collaboration which purports to test the predictions of several theories of consciousness in an open and fair way. (You can see here, from Hakwan Lau, some commentary on whether it is actually designed and executed appropriately to achieve that). The letter seems to reflect the growing exasperation of some researchers in the field with the perceived hype and misrepresentation of IIT, its claims, and the results of the COGITATE study, which apparently came to a head and prompted this drastic action.

I’m not working in this field and don’t have a dog in this fight. I don’t know whether publishing such a letter was warranted or will turn out to be helpful or damaging to the field. I will say it’s not a good look. But then neither is the hype or the indulgence of theories that are not actually scientific. My own feeling is that it’s not so easy to say what is and what isn’t scientific. When it comes to IIT, I think some aspects of it are specific enough to be testable (or at least people think they are), while some lead to metaphysical positions I personally find absurd, such as panpsychism, which are not testable and not, in fact, scientific, in the operational sense (nor are they intended to be).

Anyway, all of that is preamble. What I really want to say here is that we do not have any “theories of consciousness”. We have a bunch of different theories relating to different aspects of consciousness (very specific aspects, in some cases). From the outside, it seems to me that many of the supposed disagreements in the field arise because these theories are really talking about different things.

For those keeping track, these include the aforementioned IIT, the Global Neuronal Workspace Theory, Higher-Order Thought theories, the Attention Schema Theory, Drive Theory, Unlimited Associative Learning Theory, Beast Machine Theory, and many others (reviewed nicely here). People in this field do seem to like naming their ideas!

Even in the very limited domain of visual perception (which hugely dominates the field), different theories may be related to different sub-questions. Why does only some sensory information reach consciousness? How is that information bound together? Where is the information actually held? Does it all pass to some higher-order areas or are the details maintained in lower levels?

These kinds of questions have been the major areas of focus in the field – so much so that they can seem like what consciousness science is about. (Just like the field of decision-making often turns out to be just about whether monkeys can tell if a bunch of dots on a screen are moving more to the right than the left). But these questions were selected mainly because they can be operationalized and are experimentally tractable. In my view, the theories that deal with these paradigms are just not theories of consciousness more broadly.

A real grand theory of consciousness, will, in my view, have to take elements from all the current micro-theories. And then plenty more, if we are to answer the kinds of questions listed below. And, even if such a theory can’t currently answer all those questions, it should at least provide an overarching framework (i.e., what a theory really should be), in which they can be asked in a coherent way, without one question destabilising what we think we know about the answer to another one.

So, here goes – from my conscious brain to yours – a non-exhaustive list of questions that occur to me that a theory of consciousness should be able to encompass:

What kinds of things are sentient? What kinds of things is it like something to be? What is the basis of subjective experience and what kinds of things have it?

Does being sentient necessarily involve conscious awareness? Does awareness (of anything) necessarily entail self-awareness? What is required for “the lights to be on”?

What distinguishes conscious from non-conscious entities? (That is, why do some entities have the capacity for consciousness while other kinds of things do not?) Are there entities with different degrees or kinds of consciousness or a sharp boundary?

For things that have the capacity for consciousness, what distinguishes the state of consciousness from being unconscious? Is there a simple on/off switch? How is this related to arousal, attention, awareness of one’s surroundings (or general responsiveness)?

What determines what we are conscious of at any moment? Why do some neural or cognitive operations go on consciously and other subconsciously? Why/how are some kinds of information permitted access to our conscious awareness while most is excluded?

What distinguishes things that we are currently consciously aware of, from things that we could be consciously aware of if we turned our attention to them, from things that we could not be consciously aware of (that nevertheless play crucial roles in our cognition)?

Which systems are required to support conscious perception? Where is the relevant information represented? Is it all pushed into a common space or does a central system just point to more distributed representations where the details are held?

Why does consciousness feel unitary? How are our various informational streams bound together? Why do things feel like *our* experiences or *our* thoughts?

Where does our sense of selfhood come from? How is our conscious self related to other aspects of selfhood? How is this sense of self related to actually being a self?

Why do some kinds of neural activity feel like something? Why do different kinds of signals feel different from each other? Why do they feel specifically like what they feel like?

How do we become conscious of our own internal states? How much of our subjective experience arises from homeostatic control signals that necessarily have valence? If such signals entail feelings, how do we know what those feelings are about?

How does the aboutness of conscious states (or subconscious states) arise? How does the system know what such states refer to? (When the states are all the system has access to).

What is the point of conscious subjective experience? Or of a high level common space for conscious deliberation? Or of reflective capacities for metacognition? What adaptive value do these capacities have?

How does mentality arise at all? When do information processing and computation or just the flow of states through a dynamical system become elements of cognition and why are only some elements of cognition part of conscious experience?

How does conscious activity influence behavior? Does a capacity for conscious cognitive control equal “free will”? How is mental causation even supposed to work? How can the meaning of mental states constrain the activities of neural circuits?

If we had a theory that could accommodate all those elements and provide some coherent framework in which they could be related to each other – not for providing all the answers but just for asking sensible questions – well, that would be a theory of consciousness.

Reflections on “Systems – the Science of Everything”

noreply@blogger.com (Kevin Mitchell) — Mon, 22 May 2023 15:12:00 +0000

Did you ever get the feeling, when you’re working on some problem (scientific or otherwise), that there are some basic principles at play that elude you, but that must have been worked out already by somebody? That’s certainly been my experience in my career in biology, whether it was in developmental biology, human genetics, neuroscience or other areas. I’ve felt the joy of discovering new components of systems and working out some interactions and pathways, but also a nagging feeling that I was not seeing the whole picture – that I was elucidating details of what was happening, but not grasping what the system was doing. I often felt like I lacked the principled framework to even approach that question. This was not because such frameworks don’t exist but because I had never learned about them – systems principles had simply not been part of my education.

This seems to be true across many disciplines. We’re all so specialised that we have to concentrate on the specifics of our own topics, whether it’s computer science, economics, biochemistry, cognitive science, or anything else. Discipline-specific knowledge is key, of course, but across all these areas there are also higher-order, abstract systems principles at play that we can recognise and think about and teach about. We can attempt at least a science of everything – a principled, rigorous, formalisable way to approach complex systems, whether they are ecologies or economies, neural networks or social networks, organisms or organisations.

That is what my colleagues and I set out to do with a new module that we recently taught for the first time at Trinity College Dublin. It is called (a bit presumptuously) “Systems – the Science of Everything”. This is one of a diverse set of modules known as Trinity Electives, which are open to students from disciplines across the university, taken in their second or third years. My co-coordinators for the module were Prof. Harun Siljak from the School of Engineering and Prof. Mary Lee Rhodes from the Trinity Business School. Here’s what we said in our blurb about this module:

“Humanity faces a growing number of ‘wicked problems’, from pandemics, to climate change, to threats to democracy. It is clear that the solutions to these problems will not come from any one discipline, acting in isolation. Yet our educational and research systems are still aligned into traditional disciplinary silos, defined by distinct methods, jargon, perspectives, and conceptual frameworks. However, across diverse disciplines, there are often common principles at play, which can be highlighted by a focus on systems. From organisms to corporations, ecologies to economies, neural networks to social networks, there are shared dynamics, abstract mechanisms, and emergent functions that arise through the organisation of elements, whether those elements are cells, transistors, people, or corporations.

The aim of this module is to explore those principles, to recognise commonalities and deep correspondences between seemingly diverse phenomena, to introduce students to the meta-science of systems, and to promote interdisciplinary or even supradisciplinary thinking. Ideas will be presented conceptually, with examples of convergence across diverse areas, including: genetics, neuroscience, computer science, business, economics, engineering, ecology, biochemistry, physics, evolution, medicine, climatology, and sociology.”

The challenge was to develop a course that surveyed the many different theories that collectively make up “systems science” – network theory, information theory, dynamical systems theory, and so on – but at a level that was accessible to mid-career students from diverse disciplines. In particular, we could not assume any real level of mathematical or programming expertise. Our goal was not to make students expert in any of these areas but rather to expose them to these mostly unfamiliar ways of thinking and send them back to their own disciplines with a fresh perspective and an arsenal of new, powerful concepts.

That’s certainly how I’ve felt over my own career whenever I’ve learned these kinds of ideas – I felt like it made me smarter. Like I didn’t have to try and work out complicated stuff from first principles every time I encountered some new topic. Instead, I could often recognise the dynamics at play and make sense of what was going on (or at least what kind of thing was going on) with reference to these general principles.

The curriculum

We faced a big problem in developing this curriculum – not in deciding what to include, but in accepting all the things that we couldn’t include! The course was organised in two, two-hour sessions per week for eleven weeks of the semester. Our Tuesday sessions were devoted to lectures and our Thursday sessions to tutorials and group work, using a software system for agent-based modelling called NetLogo – more on that below. That meant we had only two hours to cover each topic we chose, and we wanted to do it from the diverse perspectives of the module coordinators – engineered systems, living systems, and social systems. In addition, we had brief guest contributions by colleagues from other disciplines across the university for many of these topics.

So, we couldn’t go deep – all we could really do was introduce the various ideas, show students that theories about these topics exist, and hope to engage them in deploying these ideas in the context of a group project and an individual reflective essay (more on those below too).

There is no right way to structure this kind of course, but what we tried to do was introduce the simplest ideas first and build from there. This is the curriculum we landed on:

1. Introduction and overview

2. History and philosophy of systems thinking

3. System structures

4. System functions

5. System dynamics

6. Information and Meaning

7. Self-organisation

8. Evolution and Learning

9. Decision-making

10. System failures

Here’s an outline of what each of those sessions touched on:

1. Systems intro and overview

· Systems challenges – need for systems solutions

· What is a system?

· Idea of common, abstract underlying principles – why a science of organisation is possible

· Personal reflections of module coordinators

2. History and key concepts

· Overview and timeline of systems approaches/theories

· Historical examples: General systems theory; Cybernetics/control theory; Systems thinking

· Why these approaches fizzled out (or did they?)

· Guest lecture: Philosophy of systems – entities, boundaries, levels, reduction/holism/emergence, causation

3. System structures

· We can scientifically characterise the structures of systems

· Boundaries

· Components and connections

· Networks (basics of network theory)

· Hierarchies (scalar (nested); functional)

· Markets

· Guest lecture: networks and hierarchies in an international peace-keeping mission

4. System functions

· The way components are connected can impart functionality

· Communication channels; operators

· Basic motifs / primitives (amplifier, filter, oscillator, switch…)

· Logical operations (informational causation from configuration)

· Higher-order functionality by combining primitives

5. System dynamics

· Global dynamics emerge from structure and function

· State space, energy landscapes

· Phase transitions

· Attractors, equilibria

· Chaos and criticality, non-linearity

· Stocks, flows, variables

· Simulation and visualisation

6. Information and meaning

· Organisation can have causal power

· Shannon information and entropy

· Signal transmission; codes

· Noise and uncertainty

· Mutual information, correlation, aboutness

· Functional information: semantics

· Information flow through networks

· Guest lecture: curiosity-driven learning in infants

7. Self-organisation

· Global dynamics, structure, behavior from local rules

· Cellular automata

· Synchrony; whole-part dynamics

· Autocatalytic sets

· Protein folding; tissue morphogenesis

· Embryonic development; gene regulatory networks; energy landscapes; attractor states

· Selection for persistence and robustness

· Complex responsive processes and structuration

8. Evolution and Learning

· Adaptation over different timescales

· Basic algorithm of variation plus selection

· Natural and sexual selection

· Functional design through trial and error

· Cumulative, directional change, creativity – the adjacent possible

· Convergent evolution in design space; discovery versus invention

· Emergence of purpose and normativity (value, meaning)

· Modularity, robustness and evolvability

· Guest lecture: reinforcement learning in neural networks

9. Decision-making

· Optimisation problems

· Bounded rationality

· Speed/accuracy/cost trade-offs; opportunity costs

· Constraint satisfaction over multiple goals

· Cooperation and competition

· Metacognition, uncertainty

· Creativity – search space, escaping local minima

· Behavioral control, policy-setting

· Ethical frameworks for guiding decision-making

· Guest lecture: Game theory

10. Systems failure

· Unanticipated consequences

· Failure modes and effects analysis

· Cascading failures

· Catastrophe theory, tipping points, resonance

· Robustness and fragility

· Fail-safes, redundancy, degeneracy

· Designing for failure; formal methods and invariants

· Unintended consequences

· Neuropsychiatric disorders as maladaptive attractor states

That is, admittedly, a lot. Each one of those topics could be a lecture course in itself. All we could hope to do here is expose the students to these ideas, make them aware that these ways of thinking exist, and hope that they can see how to incorporate some of the ideas into their own ways of thinking. Complex systems are complex, but not hopelessly so. We can use these abstract principles and formalised theories to “grok” them – to see what’s going on and understand why, even when the details and particulars vary.

Student engagement, group work, and assessments

This module was not designed with a traditional content delivery focus. Our goal was not to tell students a bunch of information and then ask them to tell it back to us in an exam. Instead, we wanted to engage students in thinking about the concepts and principles of systems and give them opportunities to deploy those concepts in new scenarios, including through computer simulations. There were three elements of such work:

Discussion boards: students were encouraged to engage with online discussion boards, where they were expected to respond briefly to a prompt and then respond to two other student contributions. This worked pretty well – most of the students were very engaged with lots of interesting back and forth. This was also a useful forum for the coordinators to see if the ideas were landing as we hoped.

Group projects: Our Thursday sessions were devoted to group work, with tutorials on the agent-based modelling platform NetLogo. These were taken directly from the excellent book: Agent-Based Modeling for Archaeology, by Romnowska, Wren, and Crabtree.

Given the varied background of students, our aim here was not to make them expert NetLogo programmers. Instead, we wanted to introduce them to the idea of simulation as a powerful approach to understanding complex systems, while recognising the limits and simplifying assumptions of any modelling.

Our cohort of sixty students was split into ten groups of six. At the start of the course, each group selected a project topic framed around a real-world systems problem. These ranged from strategies to foster the resurgence of red squirrels in Ireland, to tackling the spread of misinformation online, to planning the perfect pub crawl in Dublin. The assignment was to research the topic, identify systems principles at play, consider which parameters might be modelled and outcomes measured, and construct a toy model in NetLogo to simulate some of these dynamics.

Most of the students warmed to this challenge over the eleven weeks of the semester. Each group was quite varied in background, which meant it took a little while for them to figure out how to work together effectively, but that was part of the point of the exercise. It forced students to think about a problem fairly widely at first, but then narrow it down and formalise it in a precise enough way to be able to model some aspects of it in NetLogo. This would come naturally to students in some disciplines, but for others it will have been a new experience. Ultimately, all the groups presented very credible, some excellent, projects, reflecting a lot of work and real engagement.

Individual reflective essay: The final exercise for the students was to submit an essay reflecting on what they had learned in the course, and considering how systems principles apply to some topic in their own discipline or to some global challenge. These ran the gamut from climate change to supply chain management to the life cycle of stars to Bayesian game theory in Pokémon! There was a range of performance, of course, but overall the essays showed that the students really were approaching things with a fresh perspective.

Student feedback

Reflections on the group project and in the individual essays showed that most students really were getting what we hoped for from this module (some more than others, of course). But before we get too self-congratulatory, we should recognise it wasn’t all rosy! In module feedback, some students commented on the workload for a module of this size, the challenges of learning a new computing language, the difficulty in following shifting perspectives from different lecturers, unfamiliar jargon, issues with group dynamics, and some other issues. Some of this is to be expected, but we will definitely have to look at the workload and also the way we’re presenting various topics in our next run of this module. We live and learn!

How many neurons does it take to change a lightbulb?

noreply@blogger.com (Kevin Mitchell) — Wed, 14 Dec 2022 16:55:00 +0000

I’ve been reading this excellent paper by David Barack and John Krakauer, on “Two Views on the Cognitive Brain”, and it made me wonder about which mode of nervous system function might have come first. To use their terminology, the “Sherringtonian” view (named after Charles Scott Sherrington) focuses on individual neurons as the elementary units of control, computation, and cognition. In this view, neurons can be thought of as individual relays in a control circuit (such as a reflex) or as elements performing discrete logical operations, which can be combined into larger circuits to carry out more complex computations. It’s all very bottom-up, algorithmic, and mechanistic (and, indeed, provided the inspiration for artificial neuronal networks, as conceived by McCulloch and Pitts). The “Hopfieldian” view (after John Hopfield), by contrast, takes the view that more global and dynamic patterns of activity across populations of neurons are the elements encoding and representing cognitive objects.

The Sherringtonian view can be applied to simple reflex circuits and seems like the right way to describe what they are doing and how they work. (It’s certainly the traditional way). By contrast, the Hopfieldian view seems much better suited to describing the functions of larger brain regions, especially in the cerebral cortex (see figures below, from the paper). In the Hopfieldian view, the details of the firings of individual neurons are not so important – what matters are the global patterns that emerge. Barack and Krakauer propose that those kinds of dynamical population-based patterns are essential as representational vehicles for a truly cognitive internal economy of the kind we see in humans. It’s easy enough to reconcile these two views as different perspectives that apply more or less in different parts of the nervous system (meaning the human nervous system).

But a different question occurred to me in thinking about these two modes of function: which one arose first in evolution? In creatures with very simple nervous systems, do the neurons work in little Sherringtonian circuits, or did neurons in fact arise as distributed Hopfieldian networks?

It seems pretty natural, naively, to assume that population-level encoding and representation would only have evolved once nervous systems reached a certain size and complexity. That is, that nature would have started with isolated neural circuits and scaled up from there, by multiplying and combining them, resulting in some emergent properties that it then capitalised on.

And maybe that’s right. Maybe creatures with small nervous systems carry out all their information processing and other neural functions based on the activities of single neurons and the discrete transformations carried out by connections between them. And maybe the functions of each of those single neurons became functions of populations of neurons as nervous systems got bigger.

But maybe it’s completely wrong. Maybe even the simplest nervous systems actually work in Hopfieldian ways, characterised by field dynamics, with attractors and low-dimensional manifolds in collective state spaces, rather than like motherboards composed of discrete logic gates arranged in particular ways. This raises the question: how many neurons do you need to create those kinds of collective state spaces? In creatures with really simple nervous systems, do they even have enough individual elements to generate these kinds of dynamics?

That set me to wondering: what animals have the fewest neurons? A twitter query threw up lots of interesting examples. The well-studied nematode, Caenorhabditis elegans, was a popular candidate, with exactly 302 neurons in hermaphrodites (all identified, named, with developmental lineages and neuronal connections known). But all kinds of other wonderful creatures were mentioned too, several courtesy of evolutionary neuroscientist Gáspár Jékely, who knows all the weirdest and wonderfullest critters.

These include the dwarf male of the annelid polychaete species Dinophilus gyrociliatus, which has 68 neural cells, notably concentrated in two centers: its “brain” (or frontal ganglia) and its penis. This organisation apparently “accords well with the bipartite behavioral pattern, which is entirely devoted to locomotion and copulation”. Ah, the simple life!

Even simpler are the parasitic “orthonectid” annelids, which grow and divide inside various marine host species, generating male and female larvae, which are basically just motile reproductive organs. Some species, such as Rhopalura litoralis, have a dozen neurons or so, but the winners are the males of Intoshia vaiabili, which have precisely two neurons!

It doesn’t seem possible that a nervous system with only two neurons could be working by generating dynamical state spaces. Those neurons have got to be just acting as individual electrical elements, in a Sherringtonian fashion… right? You can’t have a Hopfield network with only two neurons! But how many neurons do you need to get those kinds of population dynamics? And how many did evolution start with?

Naively (I mean, really naively) you might think, since nervous systems have grown in size and complexity along many lineages, that they must have started small and simple – with only a few neural cells. Among extant species, you’ve got sponges, with no neurons, and then you’ve got things like comb jellies or cnidaria, that have neurons (quite a few, as it happens). So how many neurons did the first critter with neurons have?!?

Did evolution start with just a couple neurons and build from there? That doesn’t seem to have been the case. Examples of animals we see today with very few neurons, like those mentioned above, may actually have reduced their numbers over evolution – perhaps due to adopting a parasitic lifestyle or due to other ecological reasons that reduced the need for a complex nervous system (like their only behavior being copulation!). It seems more likely that evolution started with lots of proto-neurons, which then evolved into lots of neurons.

There are a number of somewhat competing (maybe really complementary) hypotheses about when and where neurons arose, and what problems they solved for early multicellular animals. On a cellular level, most of the components that neurons use to do their jobs were already present and being used in other types of cells. This includes the machinery for regulating electrical potential across the membrane (ion channels and transporters), for coupling electrically to neighboring cells (gap junction proteins), or for communicating with other cells chemically (secretory systems and a diverse range of receptors). The real specialisation of neurons may have been their morphology – their long, branching projections that allow them to contact cells far away, bypassing intervening cells, and to connect to multiple cells at a time. In addition, their polarity – having an input side and an output side – allows them to propagate signals in a directional fashion.

These make them ideal for solving one of the main problems that arises in multicellular animals – the need to coordinate movements of all of their parts. Many simple animals have sheets of myoepithelial cells, which are electrically coupled and capable of contraction. (Not quite muscles, but muscle-like epithelia). Sponges, for example, are capable of rhythmic movements that rely on traveling waves through such sheets. What neurons give you is a more specific way to coordinate the contractions of these sheets of cells, which, in parallel, may have evolved into more discrete contractile units, i.e., muscles. One hypothesis for the early evolution of nervous systems – the “skin-brain thesis” – posits that neurons emerged as specialised cells intercalated within these myoepithelia, providing a “horizontal” coordination across the entire animal.

An alternate hypothesis focuses on chemical communication systems. Many simple animals also have secretory systems of peptides or hormones that can modulate the contraction of these sheets of cells. The release of such chemical signals is typically not very localised, however. Again, neurons provide a solution – the ability to target release of chemical signals at very specific sites in the organism (along with the ability to selectively control responsiveness through differential expression of receptor proteins). The “chemical-brain hypothesis” proposes this kind of localised chemical signaling as the earliest function of nervous systems.

This figure and the one above reproduced from: Arendt, Detlev (2021)

Elementary nervous systems Phil. Trans. R. Soc. B3762020034720200347

There are arguments in favour of both of these hypotheses and perhaps both functions emerged in a coordinated fashion. In both cases, it’s notable that what is posited is a “horizontal” coordinating function, rather than a “vertical” connecting function that links sensory inputs to motor outputs in some specific, reflex-like way. Neurons obviously offer the potential to make such functional couplings between particular stimuli and coordinated behavioural responses, once they are in place, but that may have been a secondarily evolving role.

What’s important for our discussion is this: in both these models, neurons emerged as new elements in a pre-existing network of non-neuronal cells. Those cells were already interconnected, often both electrically and chemically – comprising fields or populations with emergent dynamics. It’s not that neurons were added to these networks from outside – it’s more likely that some of those cells became neurons. The earliest neurons may thus have been born from and into Hopfieldian collectives.

Barack and Krakauer specifically propose population patterns as vehicles representing cognitive objects – elements they argue are necessary to explain high-level human cognition. (An argument I agree with). But perhaps the Hopfieldian model is just how neurons work, more generally. Maybe it’s how multicellular organisms work – as fields of cells, rather than discrete, machine-like components. Perhaps the Sherringtonian model, if it applies at all, reflects a derived function, an exception to this more general pattern. Or maybe it’s just an artefact of a forced perspective – an illusion of specificity created by an experimentally isolated system.

We teach neuroscience using the reflex circuit as a tidy, well-behaved, easily understandable exemplar to illustrate neural circuit function. This is potentially misleading in a number of ways. First, it emphasises an input-output function and gives an erroneous view of the nervous system as a passive, stimulus-response machine. Second, as pointed out by John Dewey, focusing on the unidirectional reflex ignores the crucial return part of the “circuit” – the action of the organism that changes the nature of the stimulus it’s receiving. And third, starting with the reflex can give the impression that such circuits are somehow the building blocks of the whole nervous system, or at least the most primitive instantiation of what can become more complex kinds of circuits. But there’s no indication that reflex circuits represent a primitive “kernel” from an evolutionary point of view.

It’s striking that the Sherringtonian approach has not succeeded in revealing the logic of the workings of even simple nervous systems or isolated subsystems. For example, the full connectome of the C. elegans nervous system has been known for decades, and all kinds of powerful tools have been applied to studying how the nervous system mediates and governs the small repertoire of behaviors that these animals display. And yet, even the simplest behaviors resist any reduction to a few discrete circuit elements, as discussed here by Cori Bargmann and Eve Marder. The capability to record activity from all the neurons in the animal, during awake behavior, is likely to change that, but this involves a move to a more global dynamical systems perspective, describing trajectories through state space, rather than discrete logical computations.

The same can be said for the well-studied lobster stomatogastric ganglion, which is clearly best thought of as a dynamical system with various possible operating regimes. And there are multiple other examples, from hydra to leeches to fruit flies to zebrafish, where these kinds of population-based, dynamical systems approaches are providing crucial insights. (All echoing pioneering work by Walter Freeman and much more recent perspectives, such as Luiz Pessoa’s new book, The Entangled Brain).

In a way, the debate between the perspectives of Sherrington and Hopfield parallels the earlier debate between Santiago Ramon y Cajal and Camillo Golgi about the nature of the nervous system. Golgi argued that neurons comprise a continuous reticulum, while Cajal contended that individual neurons were physically separated from one another and should be considered independent units. Cajal was right, of course, on an anatomical level – neurons really are discrete cellular units, rather than a cytoplasmically connected reticulum or syncytium. But perhaps Golgi was closer to the mark, from a functional sense. Maybe no neuron is an island – maybe, right from the get-go, evolutionarily speaking, neurons functioned as members of populations, fields of cells with shared dynamics, collectively traversing common state spaces.

Getting to the bottom of reductionism – is it all just physics in the end?

noreply@blogger.com (Kevin Mitchell) — Wed, 17 Aug 2022 09:23:00 +0000

There was some interesting recent discussion on Twitter regarding claims made in a new book by physicist Sabine Hossenfelder, in which she at least seems to assert that everything that happens in the universe is reducible to, and deducible from, the low-level laws of physics. Strikingly, she presents this view as an irrefutable scientific fact, rather than an arguable philosophical position. It’s worth digging into these ideas to probe the notion that the behavior of all complicated things, including living organisms, just comes down to physics in the end.

Patrick Baud quoted several passages from the book, “Existential Physics”, that argue that reductionist theories are the only game in town. With the important caveat that I have not read the book in full, and granting that some additional nuance is probably added elsewhere, it is worth quoting these passages in full to try and get the gist of these arguments. I’ve interspersed a few brief comments between the quoted sections:

“Countless experiments have confirmed for millennia that things are made of smaller things, and if you know what the small things do, you can tell what the large things do. There’s not a single known exception to this rule. There is not even a consistent theory for such an exception.”

This can be taken in several ways. If it just means that, if, at any given moment, you know what all the bits of a system are doing, then you know what the complete system is doing, then it’s trivial. (It is in fact a statement about us, and what we know, not really about the system, per se). If, however, it’s implying that all the causes of the behavior of a given system originate in the laws governing the smallest elements (i.e., bottom-up), then it’s a much bolder claim, one which she seems to be making here:

“A lot of people seem to think it is merely a philosophical stance that the behavior of composite object (for example, you) is determined by the behavior of its constituents – that is, subatomic particles. They call it reductionism or materialism or, sometimes, physicalism, as if giving it a name that ends in -ism will somehow make it disappear. But reductionism – according to which the behavior of an object can be deduced from (“reduced to” as the philosophers would say) the properties, behavior, and interactions of the object’s constituents – is not a philosophy. It’s one of the best established facts about nature.”

First, it’s very much a philosophical position, and a highly debatable one, as I hope what follows will illustrate. Second, reductionism should not be equated with materialism or physicalism – there are non-reductive forms of both those stances. They basically just say: “no magic!” They don’t say anything about the nature of causation in physical systems and certainly don’t rule out whole-part or top-down causation. Third, “reduced to” and “deduced from” are equated here, but in reality are not the same thing. Reducing something to the behaviour of its parts is an explanation. Deducing a behaviour from those parts is a prediction, which is often much harder, and which, in a sense, includes explaining how it came to be, not just it’s behavior at a given moment.

And finally, the phrase “and interactions of the object’s constituents” is doing a lot of heavy lifting here – it’s what governs those interactions that is at issue. You can obviously say that a system behaves the way it does, at any given moment, because its particles are arranged in a certain organisation (and the laws of physics then entail its behavior). But you can just as well say that the particles are arranged that way (i.e., they got arranged that way) because they create a system that behaves that way. This is answering a diachronic “how come?” question, rather than a synchronic “how?” question. These perspectives are complementary, not conflicting.

Hossenfelder continues:

“We certainly know of many things that we cannot currently predict, for our mathematical skills and computational tools are limited. The average human brain, for example, contains about 1000 trillion trillion atoms. Even with today’s most powerful supercomputers, no one can calculate just how all these atoms interact to create conscious thought. But we also have no reason to think it is not possible. For all we currently know if we had a big enough computer, nothing will prevent us from simulating a brain item by item.”

There’s a revealing assumption here – that conscious thought is “created by the interaction of atoms”. Is that the right way to think about it? I mean, atoms are certainly involved – any kind of physical structure with dynamics which support consciousness must be composed of atoms. But is that the right level to look for what makes those structures or those dynamics special? If you start with that position, then you’re making a circular argument.

“In contrast, assuming that composite systems – brains, society, the universe as a whole – display any kind of behavior that is not derived from the behavior of their constituents is unnecessary. No evidence calls for it. It is as unnecessary as the hypothesis of God. Not wrong, but ascientific.”

Well, this is throwing down the gauntlet! It casts any kind of holistic, non-reductive thinking as mystical – on the level of religious superstition. This will be news to all the chemists, biologists, psychologists, sociologists and practitioners of all the other “special sciences”. Physics rules, and that’s that. All the action is really at the bottom.

It’s not, however, entirely clear what “derived from” means here. Again, if it just means that a full description of the behavior of the particles of a system entails a full description of the behavior of the whole system, well, no one’s going to argue with that. If you know what all the atoms in my car are doing, you know what the car is doing. If, however, it is meant to imply that such a description provides an explanation of why the system is the way it is, how it came to be, or why it behaves that way, well then I, for one, certainly will argue. Your detailed picture of all those atoms of my car (and me, as the driver) won’t tell you that I’m driving to pick up milk.

Just to make clear that Hossenfelder is not an outlier in these positions, here’s a more explicit statement from Ethan Siegel:

“…the fundamental laws that govern the smallest constituents of matter and energy, when applied to the Universe over long enough cosmic timescales, can explain everything that will ever emerge. This means that the formation of literally everything in our Universe, from atomic nuclei to atoms to simple molecules to complex molecules to life to intelligence to consciousness and beyond, can all be understood as something that emerges directly from the fundamental laws underpinning reality, with no additional laws, forces, or interactions required.”

That really makes it quite explicit. Not only can the laws of physics help us understand why a system behaves the way it does, they can explain “the formation” of such systems – i.e., how they came to be the way they are. That may be true for things like atoms and planets and galaxies, but we’ll see below that it falls short for more complex entities, including living organisms.

Siegel goes on to say that: “The alternative proposition is emergence, which states that qualitatively novel properties are found in more complex systems that can never, even in principle, be derived or computed from fundamental laws, principles, and entities.”

This is a common contrast to draw – reductionism versus emergence. But emergence is a famously slippery and contentious concept, so much so that Siegel (like Hossenfelder) thinks any such explanations contravene not just reductionism but physicalism and amount to appeals to the supernatural or divine. Again like Hossenfelder, he takes them to be anti-scientific. (For a counter, see this discussion on emergence).

A more apt contrast to draw is between reductionism and holism. Regrettably, “holism” seems to carry some mystical connotations for some scientists as well, possibly due to its centrality in some Eastern religious philosophies and its cachet among New Age woo-merchants. But, in scientific terms, this is really a contrast between completely bottom-up causation and the (very much non-supernatural) idea that the organisation of the whole may entail constraints that collectively govern the behavior of the constituents. Much more on that below.

Physicist Sean Carroll has similarly argued that the forces and equations of the Standard Model (or Core Theory) of quantum physics are “causally comprehensive” (at least within the ranges of normal experience – i.e., not near a black hole or the speed of light). He grants – in what he calls in his book The Big Picture “poetic naturalism” – that it’s convenient to talk about things at other levels, but the implication is that this is at best a useful fiction.

Anyway, I was going to respond on Twitter with some thoughts, but they got so long I decided to list them here as a blog instead (but still in tweetstorm/bullet point format). Some (perhaps many) of these points are arguable, but I think they make a good case for the importance of higher-order causation in understanding many aspects of the universe, including our own existence.

Let’s consider some reductionist claims:

Claim 1: If we know the detailed microstates of all the particles in a system at a given moment, then we know all the information about the macrostate of the whole system. This is trivial and obvious.

[Though Philip Ball notes that even this is not universal! It’s not true of entangled states, in which information about the whole state is not reducible to information about its component particles.]

Claim 2: If we know the detailed microstates of all the particles in a system at a given moment, AND we can submit them to the equations of the Standard Model, then we can predict what the next state will be.

This is a MUCH stronger claim, and it appears to be false, given the fundamental indeterminacy at quantum levels.

We can only – in principle, not just in practice – make a statistical or probabilistic prediction of the possible subsequent states of the system. We can predict these probabilities very, very accurately, but in practice, so far, only for systems of very limited scale and complexity.

But we can’t predict the actual outcome of any given “run” or observation or measurement – only what the distribution of many such measurements would be if it were possible to make them and remake them from the identical starting position.

So, the claim that conditions at any given moment, plus the “laws of physics”, fully determine the state of a system at the next moment (and arbitrarily far into the future) appears to be false.

Claim 3: It could still be claimed that for any given system, all the important causal interactions happen at the lowest levels, even if some are random. It’s all physical forces playing out between particles or quantum fields.

Note that such a view has no historicity – it doesn’t matter how a system arrived at a certain state; all that matters is what that state is.

Systems like this have no memory – they reflect a certain arbitrary path that has been followed but they don’t accumulate any complexity. They can’t do anything and they’re not for anything.

Assumption: Note how the reductionist position already assumes that “systems” exist – collections of particles that have some integrity as an entity and autonomy from the environment, often with some internal structure.

But how would such systems even come into existence in a causally reductionist universe, without any historicity? The equations of the Standard Model, by themselves, can’t explain this tendency.

However, once there is some randomness, which creates a possibility space, statistical principles will come into play across collections of particles, over time. The laws of thermodynamics will favour some arrangements over others.

Some arrangements will be more stable than others – some will dissipate more free energy than others (or maximise the rate of entropy maximisation in the universe as a whole, even while local entropy decreases). Things will tend to get organised.

Now a simple mathematical principle will apply: more stable arrangements will persist longer. Note that what matters for stability is the global patterns of matter and forces and the patterns of flux – i.e., the dynamics of the macrostates, not the particulars of the microstates.

Claim 4: A reductionist might counter that any macrostate must supervene on some particular microstate. So it all comes down to the details at any given moment!

In fact, this relationship only goes one way – a given microstate must correspond to a given macrostate. But, for the purposes of determining thermodynamic stability, a given macrostate may be realised by multiple possible microstates.

That is, the universe itself does coarse-graining – it’s not just something we do for convenience.

If you want to claim this is all also “just physics”, well, fine, but it’s not just bottom-up causation derived from quantum theory. And we’ll see below how it leads to complex chemistry and ultimately to biology.

Claim 5: Another (very strong!) reductionist claim is that once we know the laws governing the behavior of subatomic particles, all other theories or principles can be derived from it. This position admits that higher-order principles exist but claims they are not fundamental.

This appears to not be the case. For example, Hamilton’s principle of least action or Jaynes’ principle of maximum entropy or the Free Energy Principle cannot be derived from the Standard Model. (At least as far as I know!)

There are all kinds of other systems principles and dynamics – familiar to engineers, economists, computer scientists, biochemists, evolutionary biologists – that do not derive from the laws of physics. They just hold.

The algorithm of natural selection is a pertinent example – iterations of mutation and selection will allow change to accumulate, complexity to increase, and functionality or adaptedness to emerge. This is independent of the physical medium (and, indeed, the algorithm is applied in all kinds of areas).

Principles like these apply on the global scale – to the organisation of systems. And they appear to be every bit as fundamental as the equations of quantum field theory.

Crucially, they do allow for historicity – in fact, they make it inevitable. Whatever system is favoured at any given moment becomes the initial conditions for the next moment. Because of fundamental randomness, this generates an exploratory and potentially “progressive” dynamic.

If a system happens into one specific state at time t, this may mean it can now reach a new, even more favoured state at time t+1. (Stuart Kaufmann’s “adjacent possible”)

This dynamic can lead to the amplification of very small fluctuations. Indeed, such random fluctuations are necessary to break symmetry (e.g., of the cosmic inflation) and allow inhomogeneities to emerge (like the formation of galaxies, stars, and planets).

The universe will thus become structured. Higher-order entities will arise, with a tendency to persist through time. (This is tautological but not trivial).

This opens the door to higher-order causation. But what does this mean? If we think of whole-part causation, this normally means the whole comprises some organisation of the parts, which collectively constrain each other.

Claim 6: In a sense, this just means there’s a solution to the global problem of all the force vectors – an energy minimum or at least transiently stable organisation of the system. A reductionist could claim this will just emerge bottom-up.

And of course this is true – *given the organisation at time t*. But in a universe with a real possibility space, some organisations will have been more likely to exist at time t.

So the probability of any given state at time t+1 depends on the prior probabilities of all the possible states that could have existed at time t. The path is historical.

Given that the universe is expanding, the space of possible states is too – faster than the physical stuff can equilibrate. That is, the maximum possible entropy of the universe is increasing all the time, but the actual entropy lags behind.

This means that the absolute amount of information (i.e., structure or local order) in the universe can increase, even while the absolute entropy is also increasing, because the possible entropy is increasing even faster!

Now we get to another crucial factor, hinted at above – feedback or selection, through time. Not instantaneous whole-part relations, which suggest a kind of circular causation, but diachronic relations that extend through time – a spiral, not a circle.

Any given configuration that is stable enough to persist for some time generates a new adjacent possibility space, which may allow the system to reach an even more stable state, and on and on. Now we can see the kind of dynamic that leads to the emergence of increasingly complex dissipative systems.

Here, the structure of the whole system (“inherited” from time t) creates the initial boundary conditions that shape the possibility space at time t+1 – a whole-part causation that is not logically circular due to it being diachronic.

The organisation of the system imposes constraints on its components. These constraints are every bit as causal as the physical forces at play – that is, they contribute to governing the way the system will evolve from moment to moment.

Forces are thus not the only causes. Indeed, you won’t have any physical forces without some structure, some inhomogeneities or gradients. As Keith Farnsworth argues, using Aristotle’s terms, there are no efficient causes without formal causes.

In fact, higher-order constraints can become even more causally important than the lower-level details, due to the coarse-graining and multiple realisability at play.

This kind of complexification can lead to the emergence of life. Living systems exist far from equilibrium and perform work to stay that way.

They are therefore under selective pressure to find solutions that afford the greatest dynamic stability. (Very different from the inert stability evident in a crystal).

Particular patterns of dynamic relations (e.g., in networks of chemical reactions) can, through all kinds of feedback loops in the system, remain stable under a range of conditions.

But they tend to be precarious, especially in an environment that may itself be quite dynamic. A good way to persist is to “save the settings” that govern the kinetics of all those reactions in a chemically inert substrate that is itself held apart from all that activity.

Then if the system is perturbed, it can re-equilibrate by reference to that stable informational resource. That is the primary function that DNA performs.

The benefit of course is that the DNA can be replicated, the cell can divide, and the structure can be recapitulated in the daughter cells, again by reference to the genetic informational resource, thus allowing reproduction.

DNA can thus act as a store of information and a substrate for natural selection. If a mutation arises that alters the system dynamics and makes it more likely to persist in whatever environments it encounters, then that mutation will be selected for. And negative selection will conversely act against mutations that impair persistence.

These two kinds of natural selection will act at the population level with a ratchet-like mechanism, meaning change can accumulate along various possible directions. Progress can be saved at each step along the way.

Causation in these systems is now inherently informational and historical – they are the way they are because of the history of their ancestors’ interactions with the environments they encountered.

And the way they are imposes constraints on the components. They are now aligned towards a purpose – persistence. This kind of causation is absent from a reductionist’s worldview.

The effects of natural selection can lead to a new kind of structure: a compartmentalised hierarchy, where different elements of the system are acting over different time-scales.

Some of these can provide top-down constraints to simultaneously manage short-term and long-term contingencies in an optimal manner.

This is yet another kind of causation, distinct from whole-part. It is actually part-part causation, but it relies on a nested, hierarchical structure, where information flows bottom-up and top-down and side-to-side, with each part trying to satisfy its own constraints, based on the context supplied by the rest of the system.

Crucially, those configurations can encode control policies – what to do in the case of some internal or environmental conditions. Optimal policies will have been selected for based on prior experience.

This is doing things for reasons. There are no reasons in a reductive world, where all the causation inheres at the lowest levels. But where exploratory dynamics lead to the emergence of true complexity, in the form of life, reasons become perfectly real causes of what happens.

Note that none of this requires “new physics” or in any way contravenes the Standard Model.

The laws describing the low-level forces remain the same. They are just subject to additional constraints in the form of initial and boundary conditions – no magic required!

But the system now crucially has historicity packed into its configuration – it means there is a why as well as a how, to how it evolves.

Ultimately, if we want to really understand things in the universe above the scale of atoms, we need to take seriously the evidence that such entities and systems behave according to higher-order principles – the particles are just the stuff they’re made of.

The accusation that these higher-order principles are somehow unscientific or mystical or appeal to supernatural forces is thus unwarranted. Indeed, to claim that the successes of particle physics in its own arena mean that every phenomenon at every scale and degree of complexity will also ultimately be explainable by these low-level laws is such an extrapolation beyond empirical evidence that it is the position that starts to look like an article of faith. :-)

On conceptual rigor: “What do we take ourselves to be doing?”

noreply@blogger.com (Kevin Mitchell) — Fri, 15 Jul 2022 15:57:00 +0000

I had the pleasure recently of attending a Festschrift celebration for Prof. Dorothy Bishop, who has been a leading light in the study of neurodevelopmental disorders (especially speech and language disorders) for decades. She has also been a vocal and effective proponent of improving reproducibility in scientific research, writing on her popular blog, on twitter, in journal articles, and even presenting in the houses of parliament on the subject.

Dorothy’s efforts, along with many others in the Open Science “movement”, have helped to highlight crucial issues of methodological and statistical rigor that have led to irreproducibility across many fields, along with potential solutions and ways that the scientific community collectively can address these problems. These efforts are aimed at improving the quality of what we are doing. But there is sometimes a deeper question to be asked: why are we doing what we are doing? What is the conceptual basis for the research we’re conducting? Of course, this is often very well laid out and supported, but not always. There is frequently a need not just for more methodological and statistical rigor, but more conceptual rigor too.

Methodological and statistical problems

Problems of irreproducibility have arisen across fields, with social psychology garnering probably the most public attention. There is really nothing unique about this field, however – these problems are extremely widespread, as pointed out already in 2005 by John Ioannidis. They arise due to under-powered studies, small actual effect sizes, poorly defined hypotheses, exploratory analyses of a high number of variables, excessive degrees of freedom in analyses including mining through covariates, lack of correction for multiple tests, and lack of internal replication – effectively fishing for significance in noisy data. These practices, compounded by publication bias, are guaranteed to flood the literature with false positives – spurious findings that do not replicate.

In the study of neurodevelopmental disorders, these problems have manifested especially in genetics and neuroimaging. The issues with candidate gene association analyses are, by now, well known. The premise here is that if some trait or condition is heritable, then there must be some underlying genetic variation associated with it. Some of that variation may be in the form of common genetic variants in the population. You may have a hypothesis about what causes a certain condition – say, derived from knowledge of the relevant pharmacology – and that may lead you to select a number of “candidate genes” for genetic analysis (such as the serotonin transporter gene for anxiety or depression, or dopamine receptor genes for schizophrenia).

The next step is to identify some common genetic variants in your set of candidate genes and perform an “association analysis” – this amounts to assessing the frequency of the different versions in a set of cases and controls and looking for significant differences. If you find a significant increase of a particular variant in cases, you can infer that it is associated with increased risk of the condition (in the same way you would for any epidemiological “exposure”).

The problem was that people searched in samples that were too small, analysing too many variants at the same time, without correcting their thresholds for statistical significance for all those tests, sometimes added in covariates like sex or environmental exposures (increasing the number of tests multiplicatively), performed inferential statistics on exploratory data, typically didn’t include an independent replication sample, and almost exclusively only published “positive” findings. The result? A decade’s worth of work that generated a literature of false positives, and a secondary literature aiming to work out the mechanisms underlying signals that were not real in the first place.

Thankfully, these problems were recognised by the human genetics community (including journals and funders), prompting a seismic shift in how genetics is carried out. The only solution was to cooperate, and very large consortia were formed to allow collection of samples from tens of thousands of people (rather than tens or hundreds). Technology was developed to assay common genetic variants across the entire genome, all in the same genome-wide association study (GWAS). Computational and statistical methods were also developed to ensure a high level of rigor: strict correction for multiple tests, with a genome-wide significance level of 5 x 10-8, control of possible confounds as much as possible (such as population stratification or batch effects from different sites), along with independent replication samples. Importantly, results for all variants are published, whether the individual association signals reach genome-wide significance or not.

The genetics community thus took the underlying problems seriously and acted to change the culture and practice of the whole field. The result has been a decade of discovery of thousands of common variants statistically robustly associated with all kinds of traits and disorders. (Whether you think such discoveries are actionable is another question – more on that below).

Neuroimaging is currently facing up to the same problems. A recent study by Scott Marek and colleagues shows that so-called brain-wide association studies (or BWAS, where thousands of neuroimaging parameters are compared across groups of cases and controls, looking for some kind of difference somewhere) require the same kind of methodological rigor as GWAS. That means much bigger samples than are currently used – in the thousands, rather than the tens. Again, this will require a cultural change in how this research is carried out – the cottage industry approach just won’t produce reliable results. (Indeed, the literature claiming to have found all kinds of imaging “biomarkers” of all kinds of psychiatric conditions is as inconsistent and irreproducible as the candidate gene association literature).

Meanwhile, some other topics of relevance to neurodevelopmental disorders seem to have not yet gotten the memo on reproducibility. This includes a lot of the work on supposed transgenerational epigenetic effects of stress or trauma, for example, or claims of casual effects of various components of the gut microbiome. The methods used in these fields suffer from all the same problems outlined above, which are guaranteed to produce more noise than signal.

The general solutions are pretty clear: bigger samples, fewer degrees of freedom in the analyses, proper corrections for multiple testing, independent samples for replication of exploratory findings before publication, and publishing both positive and negative results. All of those can be supported by the use of Registered Reports, where a research design is submitted to a journal and peer-reviewed (with an opportunity to improve the design at that stage), and where, if the design is approved, the journal agrees to publish the paper describing it regardless of how the results turn out. This provides a much more robust way of doing open and reproducible science, that aims to undercut the perverse incentives that can lead to systemic biases, especially publication bias.

Conceptual clarity

The recognition of these problems and the changes that various fields are implementing to solve them are very welcome. They’re crucial in fact, and should be a core part of the training of all scientists, as well as the upskilling of those of us who were trained before these issues came to light. But I wonder sometimes if they’re enough to ensure the science we do is not just reproducible but actually productive. What we should surely be aiming for is a truly progressive and collective deepening of our understanding of complex issues. To achieve that, we may need more conceptual clarity and, frankly, effort, than is sometimes demonstrated.

Philosophers have a phrase they often employ, when talking or writing about their work – kind of a self-narration of the processes of thought. They will often say: “What I take myself to be doing here is…”. This might be something like: I take myself to be making a claim about the ontic rather than the epistemic status of something or other (i.e., what kind of a thing something is, rather than just what we know about it). I used to find this habit a bit pretentious, even precious, but the more I encounter it, the more valuable I think it is.

Really, it’s a discipline of reflection on the activity one is engaged in. It makes you make explicit the premises and assumptions on which a line of reasoning is based. The reason to lay this out there, for all to see, is that you might be wrong – not about your conclusions, but, more fundamentally, about what it is you think you’re doing. Philosophers make these declarations precisely so that others can challenge them, but it is, more fundamentally, a useful exercise in clarifying your own thoughts to yourself.

I’d like to see more scientists adopt this habit. It would be hugely useful to lay out the premises and assumptions underlying any given experiment or piece of research. Not just the proximal ones that have prompted some specific hypothesis, but the deeper ones that underpin the approach in general – the ones that often go unstated and unexamined.

For example, if we’re proposing to carry out a GWAS of some psychiatric condition, what is it we’re hoping to find? Presumably a list of associated genetic variants, but why? What use will they be? Will they tell us about the underlying “biology of the condition”? That phrase may have very different meanings, depending on the nature of the condition. “The biology” of cancer lies at the level of proteins controlling cellular differentiation, proliferation, cell cycle control, DNA repair, and so on. These are tightly linked to the functions of implicated genes. In contrast, “the biology” of something like autism or schizophrenia manifests at the level of the highest functions of the human mind – social cognition, conscious perception, language, organised thought. Finding the genes that convey some risk for the condition is highly unlikely to immediately inform on the biology underpinning those cognitive processes and psychological phenomena.

Instead, those GWAS have pointed in general at genes involved in neural development, reinforcing the view of the symptoms as emergent phenotypes, not directly linked to the functions of the encoded proteins. Given that is the case, we might ask what we take ourselves to be doing if we’re proposing carrying out ever-larger GWAS of these conditions with bigger and bigger samples. I’m not arguing against it, just suggesting that it’s worth making explicit what is to be gained from such an exercise. The promised insights into “the biology” of the condition are unlikely to just pop out of such studies, nor will they directly produce a list of new molecular therapeutic targets, as often suggested.

More generally, we can ask: what kinds of things are our diagnostic categories? Do we take a category like autism or depression or bipolar disorder or schizophrenia as monolithic, representing a natural kind, or instead as a diagnosis of exclusion that may encompass a myriad of etiologies and pathologies? Our basic conceptions here are crucial in answering the question of what we take ourselves to be doing when, for example, we put a hundred people with autism in the MRI scanner and compare them to a hundred neurotypical people.

If, say, we’re looking for some structural brain differences between these groups (as has often been done), we should be able to explain why we might expect to find such a thing. The genetics has clearly shown us that “autism” is an umbrella term that describes an emergent cluster of symptoms at the psychological and behavioral levels, linked with extremely diverse genetic etiologies. Should we then expect some commonalities across patients at the level of brain structure, though they may have a hundred different genetic causes? Would we propose such an experiment for a category like “intellectual disability”?

Analyses of gene expression patterns or epigenetic marks in the (post mortem) brains of subjects with these conditions are similarly vaguely justified, if at all. What is the underlying premise that is being tested? That all those diverse genetic etiologies might converge on a pattern of gene expression that underlies the observed symptoms? Should we expect the symptoms to have a direct molecular underpinning like that? Or do they reflect instead emergent activity regimes of the dynamical neural systems of the brain?

Alternatively, should we expect to see a direct signature of the primary genetic disruptions in the gene expression patterns of various parts of the adult brain? Attempts to link altered gene expression profiles to the genes directly implicated by GWAS seem to imply this idea, yet this premise is not made explicit or justified in the relevant publications. And it is not obvious, under that model, why so many distinct genetic etiologies would then lead to a consistent signature across patients.

You could do these kinds of projects with all the statistical and methodological rigor that could possibly be brought to bear and still not learn anything useful, if they are not founded on clear conceptual premises.

The same is true in many areas of cellular and animal neuroscience. Again in relation to neurodevelopmental disorders, if we make a mouse model of a mutation associated with high risk of autism in humans, is this a model “of autism”? Should we expect some behavioral similarities between the phenotypes observed in the mouse and in humans with autism? It’s probably better to think of the mouse as just a model of the effects of that particular mutation, but effects at what level? On biochemical pathways, developmental outcomes, function of neural circuits or systems, emergent behavioral phenotypes?

Again, how close a correspondence should we expect between what we see in a mutant mouse at any of these levels and what we observe in humans? Given that the effects of even high-risk mutations can vary hugely across individual humans due to differing genetic backgrounds and idiosyncratic developmental variation, what should we expect from a single, arbitrary (but inbred) genetic background in the mouse? None of this is intended to argue against doing this kind of work in cellular or animal models – it is merely a call for more conceptual clarity in laying out what can and can’t be gained from investigations at any given level.

To sum up, the focus on improving statistical and methodological rigor in these and in all fields is crucial if we want to make our science robust and reproducible. But, if we don’t take the time to ask ourselves what we take ourselves to be doing, we’re likely to end up doing something other than what we think. Poorly conceptualised experiments can waste just as much time and resources as poorly executed ones.

The evolution of meaning – from pragmatic couplings to semantic representations.

noreply@blogger.com (Kevin Mitchell) — Tue, 07 Jun 2022 06:53:00 +0000

When living creatures perceive something, they’re concerned with two questions: What is it? and: What should I do about it? You might think that the machinery for answering those questions evolved in that order – like you’d have to know what something is before you can know what to do about it – but it seems likely to have been the opposite. The actions of the simplest creatures when faced with various stimuli in the world are mostly coordinated by pragmatic couplings – signals that are prescriptive rather than descriptive. But these mechanisms laid the foundation for the evolution of decoupled internal representations with true semantic content.

For living organisms to go on persisting – which, let’s face it, is their whole schtick – they have to take in energy and raw materials (food, oxygen) and use them to keep their internal economy humming. Many organisms manage this process – known as homeostasis – by staying put and letting resources come to them. The problem with this strategy is that sometimes environmental conditions change and resources dry up – this is especially true if the organism’s own activity (along with its friends) uses up resources locally. A good strategy under such circumstances is to move, and many creatures evolved various ways of doing that – swimming, sliding, oozing, crawling, rowing, wriggling, eventually walking, running, and even flying.

But where to go? And how to know?

Some creatures – like a lot of marine plankton – simply float about on ocean currents and hope for the best. Others sense a depletion of food or a build-up of toxic waste products in their current environment and simply set off in a random direction – anywhere else being better than here. But most have evolved some kinds of specific sensors to detect relevant substances or objects in the environment in order to control or inform the direction of movement.

Some of the best-studied examples of these systems are in the simplest organisms – such as the chemotaxis responses in bacteria like E. coli. These rod-shaped bacteria move about by rotating a long thread-like filament called a flagellum that extends from one end and works a bit like an outboard motor. It has an unusual mechanism of action, though – when it rotates in one direction, it recruits a bunch of other filaments on the surface and forms one big propeller, pushing the bacterium forwards. But when it rotates in the other direction, all those other filaments interfere with it and the whole bacterium just revolves on the spot.

Most of the time, E. coli will swim in a straight line for a bit, then tumble around, and then head off in a new, apparently random direction. However, when they detect some chemical substance that is a food source – like glucose, for example – they will spend more time swimming in a straight line, as long as the signal from the detection of the food source is increasing. In this way, they travel up the concentration gradient and arrive at the spot with the maximal amount of resources.

The molecular factors that mediate this behaviour include receptor proteins that protrude like antennae from the surface of the bacterium and can bind to the food substances, and internal proteins to process and transmit the signal, ultimately controlling the direction of rotation of the flagellum. This system is often described in purely mechanistic terms, as if the bacterium were a passive, stimulus-response machine.

This is deeply mistaken – the bacterium is an endogenously active agent, constantly monitoring its environment for information that it accommodates to. In isolation, the system I just described looks very linear and deterministic – as if its activation will push the bacterium one way or another. In reality, it works in a much more integrative and holistic manner. E. coli have receptors for many different kinds of chemicals and other stimuli, including potentially harmful ones. These must all be integrated to “compute” the most adaptive direction to travel. And the signals from any potential food source must be compared over adjacent time-steps for the bacterium to be able to follow a gradient. In addition, all of this machinery can be modulated by conditions like temperature and osmolarity and the overall density of other cells.

So, there is a lot of integrative processing going on even in these apparently simple actions of very simple creatures – so much that it is sometimes referred to as “basal cognition”. However, there are some crucial differences from the kind of activities typically thought of as cognitive in animals or humans.

The main difference with fancier types of cognition is in what those signals mean to the organism. And here we get into some tricky philosophical waters. We’re used to talking about information in scientific terms – it can be localised and even quantified. Meaning is a slippier concept. It’s obviously much more qualitative and subjective – more constructed by the interpreter of a signal than residing somehow in the signal itself.

Pragmatic meaning

We can say that the activation of a protein receptor molecule conveys information to the interior of the cell – information about something out in the world. But it is also information for something. The reason bacteria have these receptors is that it is adaptive for them to be able to detect these substances out in their environment and do something about it. In fact, these simple systems are configured in such a way that the organism just does something without separately apprehending what the signal is about.

The meaning to the organism is not “there’s glucose here” or even “there’s food here” and it’s not “there’s a dangerous chemical here” or even “there’s a threat here”. The meaning of these signals to the organism is simply “approach!” or “avoid!”. They are prescriptive, not descriptive. Or, if you prefer, they are imperative, not indicative. They’re not pointing at something or reporting what’s out there – they are simply demanding action.

So, in these simple cases, the organism doesn’t infer what’s out in the world and then decide what to do. Really, there’s no inference going on at all – just an adaptive response. The meaning of these signals is not (yet) in their semantic content or aboutness – for living organisms, meaning started with salience. The information is meaningful for some entity, relative to some goal or purpose.

For living creatures, that purpose is staying alive. This is the master function that anchors everything else. It emerges straightforwardly – inevitably – from the action of natural selection. Organisms that are better at persisting, persist better. Their structures come to be configured in ways that are best fitted to surviving and reproducing in their environments. This isn’t just a passive kind of persistence, either – organisms have to do thermodynamic work to stay alive, and evolution does design work to favour ones that do it best.

This adaptation includes the configuration of the systems that integrate sensory signals and control behaviour. Creatures that tend to approach food sources and tend to avoid danger will do better than ones that don’t. These stimuli therefore have value, relative to the goal of persistence. And they come to have meaning – “approach!” or “avoid!” – wired right into the biochemical configuration, through the cumulative verdict of natural selection on the types of behaviours that organisms take in response to them.

The meaning is thus not just in the “content” of the signals. The currently active signals conveyed by receptors for these various stimuli are interpreted in the context of stored control policies that are wired into the system.

That’s all well and good – those kinds of facilities enable simple creatures to navigate their world adaptively. But they are limited in scope. This is because all that processing is happening at one level, within a single, shared space over a common timeframe – everything, everywhere, all at once. You can do a lot by merging a bunch of different signals like that but you quickly run out of degrees of freedom. There is just a limit to how many cogs you can compute with in a single, fully interconnected system before they get jammed up. The solution to this problem is to introduce some intervening layers and decouple the systems of perception and action. This is what nervous systems are good for.

Big beasts with brains

When multicellular life emerged, it faced a new problem. Animals now had bodies with lots of different bits – bits that all had to be coordinated for the animal to be able to move in an effective fashion. Muscles evolved to move the bits and neurons evolved to help coordinate them – with each other and also in response to sensory information. At first, these functions may have been performed by the same cells – ones that sat in the skin, with an outer sensory part and an inner “muscular” part, capable of contracting. But at some stage, these jobs were split into sensory cells and motor cells, and then a new type of cell came along, which sat between them – neurons.

The advantage of having this intervening layer of neurons is that they could communicate across a whole field of sensors and motor actuators to coordinate them. And they could do it fast. The biochemical signaling and processing that happens in single cells is powerful and efficient, but it’s slow and not suited to communication over long distances. Neurons use electrical signals instead, which are extremely rapid, and they evolved complex shapes and long projections that could connect elements in different parts of the organism. Organisms thus evolved new, powerful systems for behavioural control and a new language for internally representing and processing information.

As nervous systems got more complex, they started to add more and more internal layers, between the sensory and motor systems. It’s important to note that there are still some systems where sensory signals are rapidly coupled to motor outputs, with a minimum of processing. For example, many insects and fish (and even some mammals) have a hard-wired escape response to certain visual stimuli, such as a big shadow passing overhead or looming towards the animal. It’s pretty obvious how such a coupled sensorimotor system would remain adaptive – it gives speed and reliability to this crucial survival response. But the real benefits of complex nervous systems come with the decoupling of perception from obligate action, which, by contrast, gives greater flexibility and integrative control.

Adding layers of neurons also increases the amount of processing that can be done on perceptual information, enabling organisms to extract higher-order information about what is out in the world. The earliest-evolving senses were smell and touch – the detection of chemical or mechanical stimuli directly in contact with the organism. Creatures that rely solely on these kinds of senses cognitively and behaviourally inhabit the here and now – they have no access to things that are far away from them, spatially or temporally, and, as a consequence, have not developed the cognitive resources to think about them or act with respect to them.

The evolution of vision and hearing changed that. These are distance senses – they’re designed to detect disturbances in the medium around an organism, whether that is the electromagnetic spectrum or the vibrations of air or water. This allows organisms to indirectly infer the presence of objects – most importantly other organisms – that are out in the world and that may be the causes of those disturbances. This is not a simple task, however – organisms have to solve the “inverse problem” and figure out the most likely causes of patterns of ambiguous stimuli.

This is where those intervening levels of neurons earn their keep. By integrating inputs from the level below, each new level extracts higher and higher-order information. In the visual system, for example, the first levels in the retina are just responding to photons of light. But the next levels are comparing across inputs, performing contrast enhancement, and feature extraction, making inferences about edges and orientations and movements. And as that information is processed in the brain itself, further inferences are drawn about objects, where they are, what they are, and how they’re moving, relative to each other and the organism itself.

All of that information is made available to the rest of the brain to inform action as appropriate, given whatever else may be happening in the world, what the current state of the organism is, what its goals are, and so on. This kind of decoupling from action thus provides a much more flexible control of behaviour, guided by richer and deeper cognition, operating over longer timescales. It pays to think about the future when you can literally see things coming from a mile away.

The patterns of neural activity now constitute internalised representations that are “detached” from obligate action. Their job is no longer imperative, but indicative, or declarative. Whereas in simple organisms, the meaning is pragmatic, in more complex organisms, it becomes truly semantic. We can think of these systems as being configured as follows:

- Simple, pragmatic couplings: If A à do X.

- Complex, semantic representations: If A à then “A” – as in, tell everybody that: “A”.

That information (that “A” is the case) can then be used to inform all kinds of further cognitive operations, inferences, and possible actions. It’s a “see something, say something” sort of scenario. The contrast between these scenarios is like the difference between a thermostat and a thermometer. A thermostat detects and in the same process acts on information about the temperature. A thermometer merely reports it (or represents it, to anyone that is bothered to look at it).

Our visual system does the same thing – it reports its inferences about what is out in the world to the rest of the brain (and to the organism as a whole). But this creates a new problem. Now that these signals are internalised, how do downstream “users” (either other neurons or brain regions or the organism itself) know what the signal or pattern is referring to? How is the semantic content anchored?

No naked representations

The meaning can’t be just given, by the pattern itself, in isolation – that’s just some neurons firing. As experimenters, we may know that a given pattern correlates with something out in the world (which we can see independently). But how do other parts of the brain, or the whole organism know that? They can’t see the thing – they only have access to the pattern of neurons firing. That pattern can’t entail its own meaning. That would be like looking up the word “car” in the dictionary and finding the definition: “car”. It can’t define itself. The meaning of the word “car” is instead given by other associated words that describe its properties and relations: “a four-wheeled road vehicle that is powered by an engine and is able to carry a small number of people”.

A similar process happens in the brain. The accumulation of such associations starts in infancy, as babies explore their world, especially by cross-calibrating between their senses. They learn that things that look like this, feel like that and taste like that and are hard and about this heavy and I can pick them up and if I drop them they make this kind of a sound. All of us gradually build up a web of knowledge of these kinds of properties and affordances of objects – latent schemas that are activated or at least accessible when we detect an example of that type of object or even when we think of it. In this way, percepts are grounded on the basis of stored concepts. And those concepts can get progressively more abstract, encompassing hierarchical categories, causal relations, and narrative sequences of events – all the information that an organism needs, in order to decide what to do in a given situation.

The semantic content of a given representation is thus embodied in a web of associations – a set of linkages and pointers to other characteristics or properties, themselves represented in latent structures across the brain. The representation is thus discrete, but its meaning is distributed. We saw with pragmatic couplings that the meaning of currently active signals is given by the context of stored control policies. With semantic representations, the meaning of any current pattern of neural activity is given by the context of stored knowledge. All of this is calibrated through experience, rigorously selected for salience, and ultimately still used to inform action, just over many more possibilities and much longer timescales.

The organism still wants to know what it should do. But now it uses decoupled internal representations (and stored knowledge) of what is out there to inform those choices.

The riddle of emergence – where do novel things come from?

noreply@blogger.com (Kevin Mitchell) — Mon, 23 May 2022 16:14:00 +0000

It’s not true, that there’s nothing new under the sun. The universe is producing novelty all the time. Galaxies, stars, and planets, where none existed before. New elements, new molecules – life itself, with the explosion of new species, and eventually new minds, capable of new thoughts. New types of things at new levels of existence – discrete entities composed of smaller entities, arranged in specific ways. New systems with new properties and new causal powers, governed by new principles. So where does all this novelty come from?

The term emergence is often used to refer to the appearance of qualitatively novel states or processes or properties that arise when things are combined in certain ways. But as with many metaphysical terms, it means different things to different people and in different contexts, and it’s not always clear what, if anything, follows from its use. In particular, it is often not stated whether “emergence” simply refers to an observed phenomenon in need of explanation, or is supposed to be an element (or even the entirety of) such an explanation.

If we take life, or example, we can say that it emerged in the universe. This is a straightforward description of what happened – it didn’t used to exist and now it does. In a slightly different way, to say that, at any moment, life as an ongoing activity of an individual organism is an emergent phenomenon also seems uncontroversial – living things are made of components that are not themselves alive, but a whole organism is. On a finer scale, it seems equally unproblematic to say that living things have a suite of emergent properties that are qualitatively different from those of their components.

These usages are fairly unproblematic, on their face at least, but they’re also not particularly enlightening. What is the term supposed to actually imply about the new properties or about the relationship between the components and the higher-level system? Is it simply that the “whole is more than the sum of its parts”? That seems trivial, or worse, nonsensical. How could it be otherwise? Almost by definition, “wholes” have properties that derive not just from the isolated properties of their parts but from the way those parts are organised (not “summed”, whatever that might mean). That’s what makes something an entity, as opposed to, say, a pile.

Then what else is the term emergent meant to imply? It hints at complexity, non-linearity, and irreducibility, at hierarchical relationships, where new levels exhibit novel kinds of behaviour, and may even exert some kind of macroscopic or whole-part or top-down causation back onto the whole system. Those properties actually sound objective enough – they’re the kinds of things that complexity and information theory often deal with and can even quantify (e.g., 1, 2, 3). But is that it? Is emergent a synonym for that kind of complexity? There seems to be an added element that warrants the usage of the term, one that hinges on something else entirely: subjective, observer-dependent properties, like unexpectedness, unpredictability, or even interestingness. It’s not just that a system has new properties – it’s that they’re surprising, maybe even impressive.

Strong and weak emergence

David Chalmers has an interesting article probing the senses of emergence, especially the supposed distinction between “strong” and “weak” forms. This hinges mainly on whether the new properties are thought to be unpredictable in principle (the strong version) or are merely unexpected or difficult for us to predict in practice (the weak version).

For Chalmers, consciousness provides the only convincing example of a truly emergent phenomenon, in the strong sense. Here, something qualitatively completely novel and entirely unpredictable emerges from systems arranged in a certain way. Subjective experience seems to pop into existence in a way that is almost magical. It’s not just that it is irreducible to the workings of the physical brain – it’s that it’s not deducible from those workings. Nothing about the physical details could predict that subjective experience should arise from them.

I’m not convinced that’s completely right, though. If you look at the trajectory of evolution of cognitive control systems, you can see how metacognition and recursive self-reference could quite naturally and adaptively evolve and be selected for. Indeed, you could make a strong case that their emergence was in fact predictable. That said, it’s still not obvious why recursive self-reference has to feel like something. So, consciousness (really subjective experience at all) still seems like our best candidate for a truly strongly emergent phenomenon (though I’d be inclined to put life and agency in that category too).

Chalmers’ discussion of weak emergence highlights the importance of the sense of unexpectedness: “To capture this, we might suggest that weakly emergent properties are interesting, non-obvious consequences of low-level properties. This still cannot be the full story, though. Every high-level physical property is a consequence of low-level properties, usually in a non-obvious fashion” …

He goes on to suggest that “weak emergence is the phenomenon wherein complex, interesting high-level function is produced as a result of combining simple low-level mechanisms in simple ways. I think this is much closer to a good definition of emergence.” … [Note a similarity here to the idea of global dynamics emerging purely from local interactions, with no top-down coordination or control].

However: “This conclusion captures the feeling that weak emergence is a ‘something for nothing’ phenomenon”…

Finally, he suggests that “an appeal to principles of design should get us the rest of the way. We design the game of Life according to certain simple principles, but complex, interesting properties leap out and surprise us.”

That “something for nothing” aspect seems to capture what people really mean when they use the term emergence. But it is also obviously the most problematic aspect of the notion. Both weak and strong emergence both suppose that the property that is taken to be emerging really did not in any sense exist before the system in which it is observed came to be arranged in that particular way.

Discovery, not invention

I think in many cases this is a mistake. The instances of supposed emergence that most catch our attention are those where it’s not just true that some new dynamics arise, but where they enable some new functionality in the system. This doesn’t have to mean that those dynamics are invented or created de novo – in many cases, especially in living systems, they are discovered. They don’t just emerge from the lower level organisation. Quite the opposite, in fact – they often reflect systems principles that simply hold in the abstract and that constrain the organisation of the lower level components (because they confer some adaptive functionality that becomes selected for).

Evolution is in the design game. But it produces functional designs by exploration and selection, not by raw creation from the void. It’s not just luck that living systems have the functionalities they do. The principles that enable them exist and can be found by evolution.

For example, there are certain configurations of elements and relations that will constitute a filter, an amplifier, an oscillator, a clock, an evidence accumulator, a coincidence detector, a logic gate; that will execute gain control, or divisive normalisation, or stochastic resonance; that will extract principal components, or drive a low-dimensional manifold, or perform Bayesian inference. Those motifs are seen in gene regulatory circuits, in biochemical signal transduction pathways, and in neuronal circuits and systems. And of course they’re seen in our artificial designs as well. The medium doesn’t matter – the functionality is abstract.

The principles at play are the subject of fields like cybernetics and control theory and information theory, as well as dynamical systems theories that encompass the dynamics observed in living systems, of self-organisation, attractors, criticality, phase transitions, and so on. It’s interesting to ask, on a meta level, whether these theories are built of elements that are fundamental – Platonic truths that simply hold true – or are themselves emergent, in the weak sense that they can be derived from theories about simpler things.

But whether the underlying principles are ultimately fundamental or not, we can still say that many of the functional properties we observe as we go from level to level in living systems are not so new after all. They’re common in fact – shared across many kinds of systems at many different levels. It’s not a matter of getting “something for nothing” – it’s a matter of finding something useful and hanging on to it.

In these cases, it’s not obvious that the term emergence really applies or offers any useful insights. The crucial difference with, say, the complex and surprising dynamics observed in cellular automata, is that the configurations and dynamics observed in living things are for something. They’re not arbitrary. They’ve been selected for robust functionality. While they may still be astonishing, they’re not unpredictable in the same sense. There’s an understandable logic to why they are the way they are. They’ve been through the filter of natural selection, which is in the business of making predictions – predictions that the world will continue to be more or less like it was in the past, and that functionalities that promoted survival will continue to be adaptive in the future.

Living organisms are made of many such subsystems, collectively configured to promote the survival and reproduction of the organism. And the dynamics we observe at the level of higher systems (composed of many subsystems) are equally relatable to the same kinds of abstract functional design principles. Evolution has explored that space of functionalities at each level and selected the ones that confer adaptive advantages. There is scope in this process for real novelties to arise, through new combinations of these functional primitives, but this still comes about through a process of exploration and selection, rather than naïve emergence. The novel functions don’t just pop into existence spontaneously or come for free – evolution has to do some design work to find and retain them.

Emergent dysfunction

To put a little twist in the tail, there is I think a paradoxical corollary to this way of thinking. While the functionalities embodied in the configurations of living systems have been selected for and are thus not arbitrary or unanticipated, patterns of dysfunction may well be. Natural selection will favour configurations that work and that confer some selective advantage. Along some lineages, that can lead to complexification of systems, notably the nervous system. The new selective advantages that come with increasingly sophisticated cognition may be strongly selected for, but the downside is that as systems get more complex, they may develop more ways to fail. Selection for robustness may not keep pace with selection for new or improved function.

The result, in our own species, may be a propensity for a minor proportion of individuals to develop emergent states of brain function that we recognise as psychopathology. Many psychiatric conditions are quite highly heritable, and seem to reflect the impact of genetic variants on the program of brain development. The striking aspect of many of the emergent pathological states is their qualitative novelty. We might simply have expected such insults to lead to either general or more specific decrements of cognitive functions. Instead, we get novel states like psychosis or mania or obsessive-compulsive disorder – not just less function, but a shift to a different regime altogether. In my book, such states – ones that are unanticipated outcomes of insults to complex systems selected for other functions – really do warrant being called emergent. (In line with the literature on “emergent failure modes” in systems engineering).

And in this case, thinking of pathological states in that way has important implications for how we think about the (very indirect and non-specific) relationship between gene functions and psychiatric symptoms.

Go big or stay home! Small neuroimaging association studies just generate noise.

noreply@blogger.com (Kevin Mitchell) — Wed, 16 Mar 2022 21:20:00 +0000

Figuring out the neural basis of differences between individuals or groups in all kinds of psychological traits or psychiatric conditions is a major goal of modern neuroscience. In humans, investigating this has necessarily relied on non-invasive tools like functional or structural magnetic resonance imaging (MRI). Many thousands of studies have been published following a similar design: measure some functional or structural neuroimaging parameters across the whole brain and compare them across individuals or groups to look for ones that are statistically associated with variation in some psychological trait, performance on a cognitive task, or membership of one or other group. Most of these studies have sample sizes in the tens or at best the low hundreds. A new study by Scott Marek and colleagues shows convincingly that those sample sizes are at least one or two orders of magnitude too low to produce reliable results.

This shouldn’t really be news to anyone who’s been paying attention either to the field of neuroimaging or to wider developments around the reproducibility and robustness of scientific research. First, let me make clear that there are lots of neuroimaging studies that don’t suffer from this problem. These include, for example, studies investigating brain activation patterns when people are performing various kinds of tasks. These don’t need huge samples of people because they use tightly controlled experimental set-ups and usually run hundreds of trials in each individual.

It is the exploratory studies looking for brain correlates of individual or group differences in some phenotype that are susceptible to the problems that come with low statistical power and the aligned issues of excess researcher degrees of freedom and publication bias. The existence of these problems is perfectly evident just from surveying the published literature. For example, there are many hundreds of papers published claiming to have found a neuroimaging “biomarker” that is associated with any one of numerous psychiatric conditions (with depression, schizophrenia, and autism leading the list). None of these have held up. Similarly, there is no shortage of reported associations of brain imaging parameters with various personality traits of one kind or another – again, none of these appears to be robust or reproducible. The empirical observation therefore is that this literature is not producing real findings.

The reasons why are now painfully clear. I say painful because the field of genetics has been through this process already, more than a decade ago. The problem arises in performing exploratory studies looking for association with any one of thousands of parameters in samples of only dozens of individuals. It’s clearly a recipe for finding associations that are merely statistical blips. If you add in a high degree of flexibility in the way the data are analysed and filter the results with a hefty dose of publication bias (which selects for “positive” findings), what you get is a literature that is hopelessly polluted by spurious results.

This accurately characterises the literature of “candidate gene association studies”, with hundreds of papers claiming association between some genetic variants and some phenotype, based on very small sample sizes. By the mid-2000’s, the field had started to recognise that these reported associations from small studies were not robust. Very large consortia were formed to pool samples, especially of patients with various genetic conditions, and new technologies emerged that enabled genome-wide association studies (GWAS) to be carried out. These were much more rigorously statistically conducted, correcting for the hundreds of thousands of parameters being tested, including essential replication samples from the get-go, and reporting all results, statistically significant and not. GWAS have been extremely successful in identifying many thousands of common genetic variants associated with all kinds of traits and conditions. They also clearly demonstrated that the previously reported candidate gene associations were spurious.

By analogy, Marek and colleagues look at BWAS – brain-wide association studies – which compare thousands of brain imaging parameters across individuals or groups to look for any that are associated with a given phenotype. By pooling thousands of samples from multiple datasets, and running exhaustive simulations of different study designs, they show convincingly that sample sizes of thousands are needed to detect realistic effect sizes. In particular, they looked at two kinds of imaging parameters – one functional: resting-state functional connectivity (RSFC), and one structural: cortical thickness – and assessed whether these measures across the brain are associated with general cognitive ability or a general measure of psychopathology.

The results are sobering. The median effect size of associations detected (the correlation value, r) was 0.01. That means that the correlation between the brain imaging parameter and the phenotype only “explained” 1% of the variance in the phenotype. They go on to show that:

“The top 1% largest of all possible brain-wide associations (around 11 million total associations) reached a |r| value greater than 0.06. … Across all univariate brain-wide associations, the largest correlation that replicated out-of-sample was |r| = 0.16.” However, they note that: “Sociodemographic covariate adjustment resulted in decreased effect sizes, especially for the strongest associations (top 1% Δr = −0.014).”

So, with samples in the thousands, these studies detected associations with very small effect sizes. The corollary is that if the real effects are so small, samples in the thousands should be needed to detect them. This means that previously published associations with samples in the tens (on average, n=25) are VERY likely to be false positives. The only “results” that could be statistically distinguished with such small samples are ones with unrealistically large effect sizes. In fact, they state that: “At n = 25, two independent population subsamples can reach the opposite conclusion about the same brain–behaviour association, solely owing to sampling variability.”

The implications are stark (and will sound harsh, but now that we know what we know there's no point pretending it's not true):

First, the published literature of such studies must be treated with a huge degree of suspicion, or ignored altogether. This won’t be easy. No one is suggesting that all those papers be retracted. This means they will sit there in the literature and likely will continue to be cited by the unwary, as candidate gene association studies often still are.

Second, researchers should stop doing these small studies. Funding agencies should stop funding them. Reviewers should stop endorsing them. Journal editors should stop agreeing to publish them. This will be, to put it mildly, an enormous culture shock. It was in genetics and it required major sociological changes on the part of all stakeholders to move past the flawed methodologies to perform this research on the scale required to produce reliable results.

But what will small labs do that don’t have the resources for such huge studies? What will students do if they can’t run experiments on practically sized samples that can be recruited over the course of a PhD? I don’t know but I would suggest something else. There’s really no sound argument (scientific, sociological, or financial) for continuing with these kinds of studies when the evidence is so clear that they don’t yield trustworthy findings. They simply waste everyone’s time and effort and resources and pollute the literature with noise.

One solution is to form the kinds of consortia seen in genetics and indeed already in operation for brain imaging. This will no doubt be required to collect samples of sufficient size to detect these very small associations with brain imaging parameters.

There is, however, a deeper question that I think should be asked. Why do we think such associations should exist? For GWAS, the rationale was extremely clear – the traits and conditions under investigation had been shown to be partly heritable. That is, some non-trivial proportion of the variance in these phenotypes is attributable to genetic variation across the population. This means that there must exist genetic variants in the population that affect these phenotypes. GWAS is hunting known quarry.

The same cannot be said for brain imaging parameters. We simply don’t know that the kinds of traits that people look for associations with are indeed associated with differences in brain structure or function that can be detected at the (really pretty crude) level of MRI. Clearly, we can’t trust the massive literature of small studies that have suggested this. But what other evidence is there in support of that hypothesis? We should expect something to be different in the brains of people with autism or schizophrenia or depression versus people without and similarly there must be some neural differences between people who are more or less neurotic or intelligent or aggressive.

But should we expect those differences to manifest at the macroscale of neuroimaging – at the level of the size of different brain regions or the thickness of bits of the cortex or the structural or functional connectivity between various structures? I mean, maybe they would, but there doesn’t seem to be a very good reason to expect that, as opposed to expecting much more distributed and microscale differences.

Any such differences (macro- or microscale) might also be quite idiosyncratic, given the clinical diversity of diagnostic categories and high-level nature of psychological constructs. There are many ways to end up with a diagnosis of autism or schizophrenia and many different profiles of distinct psychological facets that would manifest as high extraversion or neuroticism. Add the very high background level of individual differences in brain imaging parameters to begin with and you have to question the confidence with which imaging association studies have been pursued. There’s certainly no “ground truth” of heritability that provides a solid motivation for these kinds of studies.

Indeed, it is notable that the phenotypes examined in the study by Marek are extremely non-specific: general cognitive ability and general psychopathology. If any phenotypes are going to show consistent associations across thousands of people (non-specifically across the whole brain), it should be these. [One technical question is whether the imaging results were adjusted for total brain size, which has a known, substantial correlation with general cognitive ability].

But should the success in detecting very weak and non-specific associations with these very general phenotypes imply that similarly massive samples could find associations between more specific kinds of phenotypes and more specific brain regions? That doesn’t seem to follow, necessarily. Frankly, the premise harkens back to naïve ideas of phrenology – that different brain functions could be reliably associated with the size of different brain regions (as revealed by bumps in the overlying skull). I'll put my hand up to say I've been involved in these kinds of studies myself in the past, but my expectations have changed over time.

One final question that it seems should be answerable by people proposing to do these kinds of BWAS at the scale that would be required (with the financial resources that would be required): what would be the point? Say we find that some psychological trait or psychiatric condition is statistically (really weakly) associated across thousands of people with a tiny difference in some brain imaging measure, like the thickness of some little bit of the cortex (or multiple bits). Then what? Would such a discovery (explaining a tiny fraction of the variance in the phenotype) really tell us much about the underlying neurobiology or cognitive operations? Would it even provide an entry point for future work? Maybe it would, but it seems incumbent on such researchers to explain how any positive (really weak) associations would be followed up.

The comparison with GWAS may be useful again in this regard. I have questioned the feasibility of following up on common genetic variants that have almost negligible individual effect sizes in these studies. But at least there is a discrete and definite genetic difference there that some people have and others don’t, that can be assessed in biological assays at various levels (biochemical, cellular, physiological). The same does not hold for brain imaging differences. These would not even be “there” in individuals – they’re not something you have or you don’t – only an upward or downward trend in some continuous parameter across many people. It’s not obvious (to me at least) what you could do with that information. (But perhaps there are very good options I have not thought of).

So, if I were reviewing a grant proposing a BWAS, I would first want to make sure that it had a sample size in the thousands, well enough powered to detect a small effect size. But I’d also want to know why such an association should be expected in the first place. And, if any such association were found, how that information would really advance our understanding of the neural basis of the psychological functions or conditions in question.

What have we learned from psychiatric genetics? The view from 2022.

noreply@blogger.com (Kevin Mitchell) — Mon, 07 Mar 2022 12:58:00 +0000

It has been recognised for millennia that risk of mental illness (broadly defined) tends to run in families. Modern science has confirmed that psychiatric disorders of all kinds are highly heritable – that is, the majority of the variation we see across the population in who is at risk of developing these conditions is genetic in origin. However, these conditions are not inherited in a simple “Mendelian” fashion, with clearly segregating risk, like cystic fibrosis or sickle-cell anemia. Instead, their inheritance is “complex”, which means that many genetic variants are at play, along with non-genetic factors. In addition, despite clear familial risks in general, many individual cases are “sporadic”, with no affected relatives. The last decade has seen tremendous efforts by many hundreds of scientists across the globe aimed at identifying the genetic risk factors and better understanding the etiology of psychiatric conditions.

The hope is that elucidating the genetics of these conditions will reveal the underlying biology and provide a pathway to develop better diagnostics, new therapeutics, and ultimately a personalised approach to treatment. Given how much progress has been made in identifying genetic risk variants, it’s worth taking a moment to see what we’ve learned. In particular, I want to look at how our general conception of the genetics and biology of these conditions has changed over the past two decades.

First, the view from 2000

In the early 2000’s, the prevailing view was that the genetics of conditions like autism or schizophrenia was complex (though it wasn’t at all obvious in what way). It was known that there were rare genetic conditions that could result in the symptoms of psychiatric disorders – like Fragile X syndrome or Rett syndrome in the case of autism, or velocardiofacial syndrome in the case of schizophrenia. However, these were deemed to be exceptional cases – the main pool of “idiopathic” cases (the ones with no known genetic or organic cause at the time) were considered by many to reflect the “real” conditions of autism or schizophrenia. This reflects a history of using these diagnostic terms only for cases where an organic cause could not be found (like syphilis, in the case of psychosis, for example). It doesn’t, to my mind, make a lot of sense to apply that logic to genetic causes, and as we will see below, the dichotomy between rare, Mendelian conditions that result in these symptoms, and the supposedly common, idiopathic conditions is a mirage that has evaporated as our knowledge of high-risk genetic factors has increased.

When geneticists wanted to find the mutations causing a particular disease, their main methods prior to the 2000’s were cytogenetics and linkage analysis. Cytogenetics involves looking down the microscope at the actual physical chromosomes of people with an illness. In rare cases, an illness may be caused by a missing or an extra chromosome (as in Down syndrome) or by a deletion or duplication of a chunk of a chromosome (though it has to be pretty big to be visible down the microscope) or by a translocation (where bits of two chromosomes get swapped with each other, often disrupting some genes). There was one notable success story from this approach – the identification of a translocation in a large Scottish kindred that broke the DNA on chromosome 1 in the middle of a gene that was named DISC1 for “Disrupted in Schizophrenia-1” (though as it happens some carriers of the translocation were diagnosed with bipolar disorder or major depression). However, mutations in this gene are extremely rare, making it at best another rare genetic cause that could be considered as separate to the much larger pool with complex inheritance.

Linkage analysis had also been exhaustively pursued for conditions like schizophrenia, bipolar disorder, and depression in particular. The typical approach was to try and find families with multiple individuals with these conditions and use genetic markers to track inheritance of the different chromosomes to try and find regions that co-segregated with the disease and thus locate a place on a chromosome where a disease-causing mutation might be found. Those approaches were unsuccessful (in contrast to many other types of genetic conditions, like cystic fibrosis or haemophilia or inherited forms of blindness or deafness). It was very difficult to find large pedigrees where mental illness cleanly segregated over multiple generations. Pooling pedigrees to give more power for the statistical analyses didn’t help, suggesting that different families did not share mutations in the same gene. By this stage, it was thus already clear that these disorders would not have the simple kind of genetics observed in other, Mendelian conditions, which had clearly causal mutations in one or a smallish number of genes.

Enter genome-wide association studies, or GWAS. The theoretical idea underlying this approach is that common diseases may be at least partly caused by common genetic variants. This was quite a departure from the typical way of thinking about genetic disease. The idea is that perhaps multiple risk variants exist in a population but only cause actual disease when enough of them are inherited at once. This would explain why mental illnesses run in families, as some families might carry more of these common variants than others, and it would also explain why linkage analyses had failed – there was no single causal variant to be found in any given family. Early models posited on the order of 10-12 such risk variants in the population for a condition like schizophrenia. The independent segregation of these variants would generate a normal distribution of burden across the population, with only those inheriting more than some threshold actually developing disease. This model allowed the supposed underlying “liability” to be treated mathematically like a quantitative trait, like height (even if all we could see was the manifestation of illness or not).

GWAS took a while to get up to speed, as sample sizes in the thousands did not produce many positive findings. This suggested that there were no common variants that had even a moderate effect on disease risk or they would have been found. (This actually fits with evolutionary expectations – many psychiatric diseases are associated with reduced numbers of offspring, meaning any variants increasing risk significantly should be strongly selected against and thus would not be expected to become common in the population). It thus became clear that models with 10-12 risk variants were not realistic – there must be hundreds or thousands of such variants involved. As GWAS studies got bigger (and then much bigger) risk variants started to be discovered.

These are sites in the genome where the DNA letter is variable in the population – sometimes an “A”, sometimes a “T”, for example. If the frequency of one of these is slightly higher in people with a condition than without, then we say it is “associated” with increased risk of the condition. Usually, the increase in risk for any given common variant is tiny – on the order of 1.05 times the baseline. As more and more of these risk variants were found, the total amount of risk explained started to grow, though it remained low in absolute terms (on the order of 3-7% of the genetic variance in risk). Nevertheless, there was an expectation that eventually the remaining common variation making up this “polygenic” risk would be found, explaining the genetic risk for the majority of cases in the idiopathic pool.

In parallel, more rare variants with much higher risk began to be found. This began with the identification of so-called copy number variants (CNVs), which are deletions or duplications of small chunks of chromosomes. These were too small to be detected by traditional cytogenetics but they could be revealed by new molecular techniques. Some sites in the genome are particularly prone to the generation of these deletions and duplications, meaning they recur with an appreciable frequency in the population.

A couple striking observations were made from these studies. First, many such variants arise “de novo” – i.e., in the generation of sperm or eggs. De novo CNVs tended to be associated with much higher risk and severity of disease than inherited ones. (Consequently, they also tend to be almost immediately selected against as people who develop severe disease tend not to reproduce). And second, the same CNVs showed up in patients with diverse conditions, including schizophrenia, bipolar disorder, autism, intellectual disability, and even epilepsy. These rare mutations thus clearly increased risk for psychiatric and neurological conditions in a non-specific fashion, with some other factors required to explain what conditions actually emerged in any individual carriers. Indeed, the same CNVs are often found (at lower frequency) in clinically unaffected individuals in the general population.

CNVs are a class of rare mutations that happen to be particularly amenable to study because they recur at specific sites in the genome. But we all also carry many rare mutations that are just changes to individual letters of the genome that occur at random during copying of the DNA. These can also result in disease but are much harder to find. However, improvements in sequencing technologies soon enabled sequencing of the “exomes” (the bits of the genome coding for proteins) or the whole genomes of many individuals at an industrial scale. These studies are currently identifying more and more rare mutations (both de novo and inherited) that confer moderate to high risk of psychiatric illness. The idiopathic pool is thus shrinking all the time as more and more individuals are found to carry some high-risk mutation.

However, that does not make inheritance in such cases simple. The risk conferred by these mutations is moderate to high (anywhere from a 3- to 30-fold increase over baseline) but not enough to explain all the genomic risk. For example, we know that if one monozygotic (identical) twin has a condition like schizophrenia that the statistical risk to the other twin is about 50%. Carrying the same genome as someone with schizophrenia thus increases risk over baseline frequency (~1%) by over 50-fold – far more than that associated with any of the individual rare, high-risk mutations. This suggests that other factors in the genome must be increasing risk in the individuals who actually end up with disease.

More recent studies are helping to identify these risk factors and elucidate the nature of their interactions. Some patients carry single de novo mutations that confer very high risk and are probably sufficient to explain the occurrence of disease by themselves. Other patients carrying mutations conveying lower statistical risk have been found to be more likely to also carry a “second hit” somewhere else in the genome and/or to have a higher polygenic burden of common risk variants. This is consistent with a unifying model whereby the polygenic risk acts as a genetic background modifier of the effects of rare, high-risk mutations. There is thus no need or reason to suggest that there are separate pools of patients accounting for these conditions – some entirely caused by rare mutations and some entirely caused by polygenic burden. Both factors are likely at play in most, if not all patients.

Sex is also an important variable. Males show significantly higher rates of neurodevelopmental disorders (such as intellectual disability, autism and schizophrenia) than females, while the reverse is true for conditions like major depression. Females with the former conditions tend to have more severe mutations or a greater genomic burden than male patients. And, when inherited, such mutations are more likely inherited from the mother than the father. This is consistent with the idea that females are better able to buffer the effects of high-risk mutations than males. This greater genomic robustness may be due to the presence of two X chromosomes versus only one in males. Males are thus exposed to all genetic variation on the X, which can generally decrease the robustness of developmental processes and the ability to buffer mutations anywhere in the genome. This is likely the reason why males across many mammalian species are more variable for all kinds of traits. It is unlikely to be the whole explanation however, as it does not explain why the male vulnerability is much higher for conditions like autism and attention-deficit hyperactivity disorder, only slightly higher for schizophrenia, and lower for depression. There are thus likely more specific neural factors at play as well, reflecting differences in the organisation of male and female brains.

With that summary of the trajectories of research in the field, we can look at some take-home messages and implications for understanding the biology of these conditions and how this may inform treatment or the search for new therapies:

1. Diagnostic categories have overlapping etiology. Both common and rare variants tend to increase risk for psychiatric illness across many diagnostic categories. This doesn’t mean the categories are not distinct types of endpoints, but does show they can have common origins.

2. The dichotomy between rare and common disorders is an illusion – the idiopathic pool is likely composed of patients with one or several rare, high-risk mutations.

3. A substantial fraction of cases, especially of more severe, earlier-onset conditions, is caused by de novo mutation (explaining sporadic cases).

4. Mutations in any of hundreds of genes can confer significantly increased risk.

5. A balance between mutation and selection explains the prevalence and persistence of these conditions.

6. The distinction between simple and complex inheritance is really a continuum – even high-risk mutations are modified by genetic background (as is the case for traditional Mendelian conditions).

7. Genetic architecture is thus both heterogeneous and polygenic, involving both rare mutations in many different genes and a more diffuse polygenic background.

8. Polygenic risk is shared across disorders and may also manifest in various cognitive or personality traits, such as intelligence or neuroticism.

9. Sex is an additional contributor to risk, probably through general effects on neurodevelopmental robustness (lower in males) and some more specific effects through unknown neural mechanisms.

10. Overall burden is key, with an architecture of more high-risk single mutations at the clinically severe end and more combinatorial interactions at the less severe end.

11. Making definitive genetic diagnoses and prognoses will be challenging, as the clinical presentations associated with any given mutation can be highly variable.

12. The genes implicated by both rare and common variants are enriched for brain (especially fetal brain) expression, for expression in neurons (of many types), and for neurodevelopmental processes (very generally). This is an important reality check. If genes expressed in spleen or pancreas or skin were turning up, we would rightly be worried that the associations might be artifactual.

13. There is, however, no real convergence on specific biochemical pathways or cellular processes or cell types or brain regions.

14. The relationship between genotypes and phenotypes thus remains stubbornly obscure. There may, in fact, be no proximal relationship at all between the molecular and cellular functions of mutated genes and the symptoms of the disorders that may emerge.

15. Divergence of outcomes even in monozygotic twins, who obviously share all their genomic risk factors, illustrates the fact that what is inherited is not a disorder, but a more general risk for psychopathology.

16. Whether that risk manifests as actual disease, and which symptoms emerge, is probabilistic, reflecting additional non-genetic factors. Given the lack of evidence for systematic environmental risk factors, this diversity of outcomes may reflect intrinsic stochastic variation in the trajectories of brain development. Chance thus plays a substantial role in determining which outcome from a wide possible range is actually realised by the processes of development in an individual.

17. The types of symptom clusters we observe (as opposed to all the unobserved ways the brain could in theory exhibit pathology) likely reflect reactive or emergent processes in brain development and regimes of function.

In summary, mutations in any of hundreds of different genes, involved in all kinds of molecular and cellular processes, and modified strongly by genetic background, can indirectly and non-specifically lead to altered trajectories of brain development that sometimes result in atypical outcomes that can manifest in diverse modes of psychopathology.

This is, to say the least, frustrating. Unlike other conditions, like cancer or autoimmune disorders, identifying psychiatric risk genes did not directly reveal the underlying biology. This is, in my opinion, because psychiatric conditions do not reflect proximal molecular biological disturbances in the way these other cellular-level conditions do. The symptoms of psychiatric conditions affect the highest functions of perception, cognition, and behavioural control – those are underpinned by distributed neural circuits and systems, not the actions of specific molecules.

We won’t, therefore, be jumping straight from GWAS or whole-genome sequencing to “druggable targets” for therapeutic development. If we want to understand social isolation or hallucinations or mania, we will need to look to neuroscience, not to molecular genetics for proximal explanations. This is, however, where genetics can provide vital entry points for experimental follow-up, especially rare mutations of large effect that can be modelled in animals or other experimental systems.

But the answers will not lie in studying any such mutations in isolation – each of these will have some idiosyncratic, even arbitrary array of biochemical, cellular, developmental, and physiological effects, some proximal, others indirect, cascading and emergent. In my opinion, making real progress will require a much longer-term meta-project to understand trajectories of phenotypic convergence and divergence across many such mutations. These emerge from properties of the developing brain, which can only be understood as an evolving dynamical system.

The project of psychiatric genetics has thus been astoundingly successful, thanks to the dedication of thousands of researchers across the world and the willing involvement of hundreds of thousands of patients and their families. But it is clearly only the first step in what will be a much longer journey to understanding the nature of mental illness.

Telling good science from bad – a user’s guide to navigating the scientific literature

noreply@blogger.com (Kevin Mitchell) — Mon, 09 Aug 2021 11:20:00 +0000

“Did you find it convincing?” That’s what one of my genetics professors used to ask us, a small group of undergraduates who blinked in response, like rabbits in the headlights. We didn’t know we were supposed to evaluate scientific papers. Who the hell were we to say if these papers – published by proper grown-up scientists in big-name scientific journals – were convincing or not? We thought published, peer-reviewed papers contained The Truth. But our professor (the always inspirational David McConnell at Trinity College Dublin) wasn’t about to let us off so easily. We learned quickly that it absolutely was our job, as fledgling scientists, to learn how to evaluate scientific papers.

This is not a responsibility that can be offloaded to peer reviewers and journal editors. It’s not that pre-publication peer review, when it works as intended, doesn’t perform a useful function. But those supposed gatekeepers of knowledge are as fallible as the primary producers of the research itself, as prone to hype and fads and influence, as steeped in the assumptions of their field, as invested in the shared paradigms and methods, and every bit as wrapped up in the sociological enterprise that is the modern reality of science.

Passing pre-publication peer review is, regrettably, no guarantee of quality or scientific rigour. Nor is publication in high-impact journals, which tend to prize novelty in both methods and results over incremental advances, which, by definition, are more likely to be robust. Individual readers of the scientific literature – researchers, clinicians, policy-makers, journalists, and especially students – should be empowered to critically evaluate scientific publications.

This is especially important now as preprints become more widely adopted, most recently in biology. The rise of preprint servers such as bioRxiv and medRxiv is a very welcome advance, busting the stranglehold of journals on access to scientific research. But it will require a change in how readers and members of the field engage with published research (and yes, posting a preprint is making the research public).

The internet has democratised scientific publishing and it can and should democratise the peer review process too. The hive-mind can take the place of a few selected reviewers and editors in judging what work is robust and reliable. Of course we’re not all technical experts in every area, but there are some general things to look out for that can help distinguish solid work from more shaky offerings. Other will have their own lists, but here is what I take to be markers of quality in scientific papers, or, conversely, red flags:

The good:

1. Based on a solid foundation of prior work

2. Clearly articulated hypothesis, with an experimental design appropriate to test it

3. Alternatively, presentation of the work as descriptive or exploratory

4. Right question asked at right level with right tools

5. Experimental controls

6. Sufficient statistical power

7. Appropriate statistical analyses

8. Converging lines of evidence

9. Internal replication

10. Conclusions supported by the data

All of which is to say that the papers I find most convincing build on previous work to generate a testable hypothesis or identify something worth exploring, design an experiment or an analysis at the right level to test it, perform the appropriate controls (whether experimental or statistical), don’t rely on a single method or results from a single experiment or analysis but draw on multiple, preferably diverse lines of evidence (which may each have their own limitations or confounds or biases, but at least they should be different ones), replicate their findings (especially effects or associations that showed up as ‘significant’ but that were not hypothesised a priori), and limit their conclusions to what is actually supported by the data.

In short, I like to get a sense that the authors went to great lengths to convince themselves of their findings and have been cautious and thoughtful in their interpretation. Conversely, these are the characteristics that make me say “nope” (sometimes out loud) and move on to the next paper:

The bad:

1. No clear rationale or hypothesis

2. Not founded on solid body of work

3. Wrong level to address the question

4. Reliance on single method or analysis

5. Statistically underpowered

6. Data-dredging: just looking through many measures for some statistically significant ‘finding’ – any effect or association in any direction

7. Covariate mining, multiplying likelihood of spurious ‘findings’

8. Failure to correct for multiple tests

9. No replication

10. High likelihood of publication bias

These are papers where it just doesn’t seem like the authors have a clear idea of what they’re asking – where the hypotheses are vague or not well justified – and where the experimental design, though it may generate data (sometimes lots and lots of data), is not well suited to actually answering any question, at least not at the level that the authors are interested in (e.g., looking to transcriptomics data to answer a systems neuroscience question, or MRI data to address a psychological question).

There may also be lots of technical experiments that are just poorly done (because many of them are frankly really hard), but it’s often difficult for non-experts to evaluate these. This is where reviewers who are real peers in the same subfield are invaluable. Dodgy stats are much more general, however, and the biggest red flag is a sample size that is simply too small to reliably detect the kind of effect the authors are interested in or are claiming is real. (The idea that if some ‘effect’ was detected with only a small, under-powered sample, it is more likely to be real, is a fallacy – such findings are far more likely to be false positives). And there are other well known ‘questionable research practices’ and biases that any reader can learn to be on the lookout for (see references below). Pre-registration – now offered by many journals – is a good protection against these dangers and a positive mark of reliability.

Here are a few examples that I have examined before: The Trouble with Epigenetics, Part 2; The Trouble with Epigenetics, Part 3 – over-fitting the noise; Grandma’s trauma – a critical appraisal of the evidence for transgenerational epigenetic inheritance in humans. They happen to be drawn from the field of transgenerational epigenetics, but the issues are very general and recognizable and discussed further here: Calibrating scientific skepticism; and here: On literature pollution and cottage-industry science.

Beyond these issues of experimental design, technical execution, and proper statistics, there are some sociological factors that tend to make me more skeptical than I would be otherwise when reading some papers:

The dodgy:

1. Part of currently sexy, trendy field

2. Appeals to other dodgy papers as support for general paradigm

3. Hype, spin, excessive claims of novelty

4. Over-interpretation or extrapolation beyond what the data support

5. Claiming to overturn mainstream thought

6. Someone selling something (or the potential for them to)

You would think that extraordinary claims would require extraordinary evidence. The reality in a lot of scientific publishing is that such claims can get published with only suggestive evidence because, if they turn out to be right, the paper will get hugely cited. (And even if they eventually fizzle or are shown not to replicate, the paper may still attract lots of attention and citations for some time). And when peer reviewers in the same field have a shared interest in the general paradigm being a thing (e.g., transgenerational epigenetics or social priming or the gut microbiome affecting our psychology), then every additional paper that gets published can be used as support for their own future publications (or grant applications).

As for selling something, sometimes this is quite overt and declared as a possible conflict of interest. But other times the work being presented is part of a long process of development of some reagent or approach that the authors hope may have clinical or commercial value in the future. There’s nothing wrong with that, in principle – developing new therapeutics is what pre-clinical work is all about, after all. And potential profit is one of the incentives that drives that work. But this incentive can lead to an exaggeration of claims of efficacy, over-extrapolation beyond the experimental system used, selective focus on supposedly supporting lines of evidence, and so on. It may be overly cynical, but in these scenarios, I set the gain on my skepticometer a little higher.

The not even wrong:

Finally, there are some papers where the experimental work and statistical analyses are all fine, but that suffer from much deeper problems in the underlying paradigm. These are ones where the conceptual foundation is vague, with unexamined and unjustified assumptions, poorly framed questions, fundamental category errors, or other theoretical or philosophical issues that mean the authors are not investigating what they think they’re investigating.

A common example of interest to me is papers that say such-and-such (a pattern of functional connectivity between some brain areas, or a profile of post mortem gene expression, or some other supposed biomarker) is the case “in autism” or “in schizophrenia” or “in depression”. These implicitly treat these psychiatric diagnostic categories as natural kinds, when we know that these labels are diagnoses of exclusion for conditions that are actually incredibly heterogeneous, both clinically and etiologically. My heuristic in evaluating these papers is to replace the offending phrase with “in intellectual disability” (which people don’t make the same error with) and see if the approach makes any sense at all.

Other examples include expecting complex behaviours (often with crucial societal and cultural factors at play) to be reducible to: differences in size of certain brain areas (e.g., Are bigger bits of brains better?; The murderous brain - can neuroimaging really distinguish murderers?; Debunking the male-female brain mosaic); levels of expression of certain genes (Epigenetics: what impact does it have on our psychology?; If genomics is the answer, what's the question? A commentary on PsychENCODE); or polygenic scores (Is your future income writtenin your DNA?).

These are examples from my own fields of research – I’m sure readers will have their own bugbears that exercise them as much as these ones do me. (A common one in psychology is assuming that effects seen in brief experiments in highly artificial lab settings will somehow translate to actual behavior in the real world).

Then there are much deeper philosophical issues affecting whole fields, for example whether a reductionist approach in biology is the right one, whether mechanistic concepts of cells or computational concepts of brains are appropriate, whether the multifarious factors contributing to complex processes can in any sense be disentangled, differences between predicting, controlling, explaining, and understanding, and so on.

Anyway, that’s my personal user’s guide to navigating the scientific literature. It may seem cynical, maybe even arrogant to make these kinds of judgments of other people’s work (though I should confess to having made and I hope learned from some of the mistakes listed above myself, both as a researcher and a reviewer). But life is short and the literature is vast and growing at a relentless pace – considering the issues listed above lets me winnow what’s worth reading in detail to a more manageable pile. Beyond that, if we want the general public and our policy-makers to “trust the science” in making decisions on issues of global impact, then we have a collective responsibility, in my view, to help curate the scientific literature.

Useful reading:

Ioannidis (2005). Why Most Published Research Findings Are False.

Button et al (2013) Power failure: why small sample size undermines the reliability of neuroscience.

Stuart Ritchie (2021): Science Fictions – Exposing Fraud, Bias, Negligence, and Hype in Science. 2021. Penguin books. (A super read and one I recommend for all undergrads starting out in biology, psychology, and related fields).

Dorothy Bishop, Oxford Reproducibility Lectures.

Dorothy Bishop (2019), Rein in the four horsemen of irreproducibility.

Björn Brembs (2019) Reliable novelty: New should not trump true.

Kevin Mitchell (2017): Neurogenomics –towards a more rigorous science.

Munafo et al. (2017) . A manifesto for reproducible science.

Tal Yarkoni (2020) The generalizability crisis.

Makin et al., 2018. Ten common statistical mistakes to watch out for when

writing or reviewing a manuscript.

Reframing the debate around free will - from the morass of moral responsibility to a grounding in biological agency

noreply@blogger.com (Kevin Mitchell) — Wed, 28 Apr 2021 19:55:00 +0000

A recent article in the Guardian very nicely summarised “the free will debate”, or at least the more visible parts of it. The debate is a live one as new findings from neuroscience or even fundamental physics are supposedly constantly and cumulatively limiting any scope for free will to possibly exist. The article presents the two most popular positions among scientists and philosophers: free will skepticism and compatibilism, which are championed by influential figures like Sam Harris and Daniel Dennett. The trouble is, they are by no means the only possible options, and, frankly, neither is convincing or satisfactory, in my view, for many reasons.

First, they often start out with definitions of free will or criteria for what would constitute it that are not just different from each other, but independently highly arguable. Second, both take determinism to be true, as a given. Free will skepticism argues that determinism rules out free will. Compatibilism argues that free will is – as the name suggests – compatible with determinism. What exactly is meant by “determinism” is not always clearly spelled out, however – there are in fact various meanings that are often conflated (see below). Usually the arguments actually rest on a reductive view of causality, as much as a deterministic one.

Moreover, the arguments are usually couched in terms of their implications for moral responsibility, and so the debate takes place on ground that is doubly treacherous. Indeed, people like Dennett argue that determinism is compatible with assigning moral responsibility to people for their “actions”, even though they actually do not make real choices or have real freedom in selecting such actions.

Anyway, you can read the article for more details of the respective arguments. The point I want to make here is that framing the issue solely in terms of these two positions misconceptualises the problem, misses whole fields of relevant literature, omits a number of other viable positions, and leads to endless debate, characterised by the two sides talking past each other. This is illustrated by a recent book – Just Deserts – by Gregg Caruso and Dan Dennett, in which they not only fail to reach agreement after a book’s worth of direct dialogue, they fail to even make clear to each other or certainly to this reader what it is they disagree on.

As a result, many scientists find these discussions not worth their time (as I get told repeatedly). Which is a shame - first, because the issues really are of fundamental importance in considering the human condition (and deeper questions about the nature of life itself). But also because they can actually be approached from a much more grounded, scientific perspective.

So, here are some suggestions – a mini-manifesto in twenty points – for ways to approach these issues that I think can let us make real progress in constructing a new conceptual framework for understanding agency and free will:

1. Separate questions of whether free will exists from possible consequences for systems of moral responsibility. The latter issue muddies the waters and leads to motivated reasoning. We shouldn’t be trying to make convoluted arguments to rescue a “kind of free will worth wanting” (i.e., one that can justify praise and blame). That is almost a theological exercise. We should be trying to understand what kind of will we have and what kind of freedom we have.

2. Don’t start with the most complex version we know of. You wouldn’t start trying to understand the principles of aeronautics by studying the space shuttle. You’d begin with trying to understand much simpler precursors and scaffolding your understanding by elaborating on basic concepts. We need to similarly ground the study of free will in the study of biological agency – how it is that living organisms can do things at all, how they make decisions and select actions, how they can be autonomous entities that merit being thought of as causes in their own right.

3. Define and dissociate the different flavours of determinism and specify the implications and challenges of each kind. Too often discussions of the “problem of determinism” for free will slip between very different meanings:

a. I’m wired a certain way, due to evolution and genetics and development and experience and though I make choices and do what I want, I don’t decide what I want. If I didn’t choose all the prior causes constraining me at any given moment, then I am not really making free choices. (Note that this kind of fatalism has actually nothing to do with physical determinism).

b. “Every event has a cause”. So what? I can be the cause. (Unless the additional claim is made that the real causes are at the lowest physical levels; i.e., not just determinism, but reductionism).

c. All that apparent choice is actually an illusion – it’s all just neural circuitry working away. Mental content is an epiphenomenon – you’re not doing anything. (Really neural reductionism seasoned with a pinch of dualism, but often presented as determinism).

d. Even the neural circuit view is illusory – there’s no real function at that level. It all comes down to physics – the atoms are gonna do what the atoms are gonna do. This is the real hard determinism where the future is pre-determined and you are not deciding anything (because possibilities don’t exist). The counterargument – that the universe is not in fact deterministic – is often dismissed by saying: well, then randomness settles the outcome and you’re still not deciding anything. (Note that this is also really reductionism, not determinism).

4. Make explicit the key problem of reductionism (thinking all the real causation happens at the lowest physical level). In particular, the deterministic idea that “every event has a cause” is only a problem for free will if you cannot be such a cause (or if the meaning inherent in the patterns of neural activity in your brain cannot be considered to have causal efficacy).

5. Consider the nature and sources of indeterminacy and its implications. There is clearly still disagreement about what current physical theories imply about the nature of reality, especially the ontological status and origins of indeterminacy, both at quantum and classical levels. But the majority of favoured theories (from very diverse angles) incorporate some indeterminacy in a way that gets resolved through interaction, marking the present as a boundary when the indefinite becomes definite. This is enough to override worries of a pre-determined future and to allow non-reductionist causation to occur.

6. Give the “rewinding the tape” thought experiment a rest. The framing of the question as: “could you have done otherwise?” at any given instant is conceptually misleading. There is no such thing as an instant. Decision-making happens over time. The real question is: do you actually have choices open to you in the present (which is not an instant of zero duration) and can you choose between them? Asking “could you have done otherwise?” is looking at that process after it’s been completed, when the possibilities have already been reduced to a definite outcome.

7. Develop a broader framework of causation (to include organisational/structural/configurational/criterial causation). The fact that the organisation of a system has causal power in determining how that system behaves (and how its components behave) is utterly commonplace, but has been banished from the reductionist worldview. There is a real science of systems that is not just “a convenient way of talking about” complex things – it really does capture why they behave as they do, in a way that even complete low-level descriptions miss.

8. Understand how freedom (at one level) relates to constraint (at a lower level). Constraint is not a bad word. Organising the components of a system towards some purpose necessarily involves restricting the degrees of freedom of those components. This is how higher-level function emerges – how a coordinated system can do things that uncoordinated components cannot. It’s why we pay football coaches so much money.

9. Develop concepts of causal insulation, coarse-graining, multiple realisability, hierarchy, and macroscopic causation. Living organisms resist the drive towards thermodynamic equilibrium partly by erecting a barrier between “them” and the outside world. This is not just a physical and chemical barrier, but a causal one. They sense what is out in the world through a veil and respond to that information (not to a transfer of energy or matter). The same thing happens between neurons or between whole areas in the brain. Causal insulation allows meaning to be extracted as information is transferred from one element to the next and allows that meaning to have causal power.

10. Operationalise purpose, function, meaning, and value as crucial causal elements. These properties do not have to be mysterious or vague – they can be precisely defined and conceptualised in scientific terms. In creatures with nervous systems, it is possible to develop a framework to understand how patterns of neural activity have casual power by virtue of what they mean.

11. Clarify the relationship between pragmatic and semantic meaning. An evolutionary view can chart the transition from direct sensorimotor couplings to uncoupled communication of signals (representations) to higher levels for more integrative decision-making. These are not mutually exclusive – the brain uses both.

12. Embrace a process view of life and self. Avoid the trap of thinking about “instantaneous” states and substances, as opposed to ongoing processes. Life and self are defined by continuity of processes over time.

13. Emphasise the historicity of living systems (lineages and individuals). Organisms literally incorporate information about the regularities of their environment into their own physical structure and use that to set criteria for future action. This means causation is spread over space and time. And this kind of historicity packs causal potential into the structure of the organism, analogous to potential energy – it takes work to put it in there and it can later be used to do work.

14. Consider the evolution of the nervous system as the development of more and more sophisticated levels of control of behaviour. The idea that the nervous system is dedicated to information processing and logical computations is a legacy of early thinking in artificial intelligence. While the nervous system can do those things, that is not what it is for. It is primarily a control system, the job of which is to define a repertoire of actions and choose between them. This control system has been elaborated over evolution to give greater and greater causal autonomy over longer and longer timeframes.

15. Avoid dualistic intuitions about the self as corresponding only to our conscious minds. Allow that: (a) our consciousness has a physical basis, and (b) much of our decision-making is done subconsciously, for very good reasons. Most of our daily behaviour is controlled by habits, heuristics, and policies that (thankfully) do not require constant conscious supervision. Deliberation and introspection are possible, however, and understanding such processes should be grounded in the science of executive function and metacognition, where goals, beliefs, and desires become objects of cognition. We can and regularly do inspect our own reasons.

16. Avoid abstract problems of instantaneous self-causation or top-down causation. It is not that the organisation of the whole system (the brain or the whole organism) at any given instant somehow simultaneously both comprises and influences the arrangement of its parts. A realistic picture of decision-making and action selection involves interaction between different parts of the nervous system, extended over time, sidestepping this metaphysical pothole.

17. Consider how consciousness supports or enhances the causal agency embedded in neural systems. What more does it get you? Are thoughts a model of one’s own cognition? Did a need for social communication of beliefs, goals, and reasons drive the capacity for cognitive introspection? Does this conscious access to our own reasons provide an opportunity to choose to reconfigure them?

18. Develop an understanding of long-term willed action as an interplay of present freedom and future constraint. Control of our behaviour extends through time. The conscious development of character over our lifetimes, active reconsideration of habits and heuristics, and adoption of goals or policies shape (inform and constrain) future action.

19. Embed this biological understanding of human agency in the context of culture and society. People’s choices are constrained by external factors in all kinds of ways. Even if we establish that human agency is real, that does not mean it can be deployed by everyone equally across all situations.

20. Only then, in my view, will we be able to relate this perspective to questions of moral responsibility. Our systems of morality are complex, evolved from biological systems of social cooperation, but socially constructed in diverse ways across cultures. Whether humans are biologically capable of free action in a general sense is relevant to questions of morality, but cannot be considered separately from all the social and cultural factors that also contribute to our moral systems.

Admittedly, that’s a lot. I’m not suggesting any one of those tasks will be simple or uncontroversial, either scientifically or philosophically. But I do think it charts out some actionable areas where a more productive framework for understanding agency and free will could be developed. The good news is there are lots of people working on lots of these elements and tons of new theoretical and empirical findings to inform this work.

Fleshing out this framework is my job in a new book I am currently writing: “AGENTS – How Life Evolved the Power to Choose”, to be published by Princeton University Press in 2022.

And you can hear more about these ideas in this recent talk I gave (virtually) at the Santa Fe Institute.

Does quantum indeterminism defeat reductionism? (Response to Coel Hellier)

noreply@blogger.com (Kevin Mitchell) — Wed, 13 Jan 2021 19:29:00 +0000

KM: I’m grateful to Coel Hellier (@colhellier) for writing a blogpost in response to one I wrote arguing that if determinism falls, it takes reductionism with it. Rather than expect people to bounce back and forth between these posts, I have pasted Coel’s entire blog and intercalated my responses, in bold, below. For clarity, I have color-coded his excerpts from my original blog in blue:

After writing a piece on the role of metaphysics in science, which was a reply to neuroscientist Kevin Mitchell, he pointed me to several of his articles including one on reductionism and determinism. I found this interesting since I hadn’t really thought about the interplay of the two concepts. Mitchell argues that if the world is intrinsically indeterministic (which I think it is), then that defeats reductionism. We likely agree on much of the science, and how the world is, but nevertheless I largely disagree with his article.

Let’s start by clarifying the concepts. Reductionism asserts that, if we knew everything about the low-level status of a system (that is, everything about the component atoms and molecules and their locations), then we would have enough information to — in principle — completely reproduce the system, such that a reproduction would exhibit the same high-level behaviour as the original system. Thus, suppose we had a Star-Trek-style transporter device that knew only about (but everything about) low-level atoms and molecules and their positions. We could use it to duplicate a leopard, and the duplicated leopard would manifest the same high-level behaviour (“stalking an antelope”) as the original, even though the transporter device knows nothing about high-level concepts such as “stalking” or “antelope”.

KM: I would describe this position simply as physicalism. It just states that if you made an exact physical duplicate of a living being, you would regenerate not just the low-level positions of all the atoms and molecules, but the high-level organization and properties as well. Of course you would – the high-level properties inhere in that organization. I don’t suppose anyone would dispute that, but there’s nothing reductionist about this assertion, as CH kind of concedes below:

As an aside, philosophers might label the concept I’ve just defined as “supervenience”, and might regard “reductionism” as a stronger thesis about translations between the high-level concepts such as “stalking” and the language of physics at the atomic level. But that type of reductionism generally doesn’t work, whereas reductionism as I’ve just defined it does seem to be how the world is, and much of science proceeds by assuming that it holds. While this version of reductionism does not imply that explanations at different levels can be translated into each other, it does imply that explanations at different levels need to be mutually consistent, and ensuring that is one of the most powerful tools of science.

KM: This stronger philosophical version of reductionism is indeed my target – the idea that the observed high-level properties and behaviors of complex systems (indeed even the existence of those systems in the first place) can in fact be reduced to and fully explained by the playing out of all the low-level interactions between the atoms and molecules. I agree with CH that this version of reductionism “doesn’t work”! But this is not some kind of straw man, as implied. It is a view that is espoused by many, especially in discussions on things like free will, and it is the version of reductionism that I argue is tightly intertwined with determinism.

Our second concept, determinism, then asserts that if we knew the entire and exact low-level description of a system at time t then we could — in principle — compute the exact state of the system at time t + 1. I don’t think the world is fully deterministic. I think that quantum mechanics tells us that there is indeterminism at the microscopic level. Thus, while we can compute, from the prior state, the probability of an atom decaying in a given time interval, we cannot (even in principle) compute the actual time of the decay. Some leading physicists disagree, and advocate for interpretations in which quantum mechanics is deterministic, so the issue is still an open question, but I suggest that indeterminism is the current majority opinion among physicists and I’ll assume it here.

KM: Some kind of indeterminism does indeed seem to be a majority view among physicists – I recorded an 80% agreement in a decent-sized but entirely unscientific Twitter poll. (Interestingly, the 80:20 split was the same among physicists and non-physicists who responded). However, there is much less agreement on where this indeterminacy comes from, how it should be interpreted, what it tells us about the nature of reality, and what its effects might be at the classical level.

This raises the question of whether indeterminism at the microscopic level propagates to indeterminism at the macrosopic level of the behaviour of leopards. The answer is likely, yes, to some extent. A thought experiment of coupling a microscopic trigger to a macroscopic device (such as the decay of an atom triggering a device that kills Schrodinger’s cat) shows that this is in-principle possible. On the other hand, using thermodynamics to compute the behaviour of steam engines (and totally ignoring quantum indeterminism) works just fine, because in such scenarios one is averaging over an Avogadro’s number of partlces and, given that Avogadro’s number is very large, that averages over all the quantum indeterminicity.

What about leopards? The leopard’s behaviour is of course the product of the state of its brain, acting on sensory information. Likely, quantum indeterminism is playing little or no role in the minute-by-minute responses of the leopard. That’s because, in order for the leopard to have evolved, its behaviour, its “leopardness”, must have been sufficiently under the control of genes, and genes influence brain structures on the developmental timescale of years. On the other hand, leopards are all individuals. While variation in leopard brains derives partially from differences in that individual’s genes, Kevin Mitchell tells us in his book Innate that development is a process involving much chance variation. Thus quantum indeterminicity at a biochemical level might be propogating into differences in how a mammal brain develops, and thence into the behaviour of individual leopards.

That’s all by way of introduction. So far I’ve just defined and expounded on the concepts “reductionism” and “determinism” (but it’s well worth doing that since discussion on these topics is bedeviled by people interpreting words differently). So let’s proceed to why I disagree with Mitchell’s account.

He writes:

For the reductionist, reality is flat. It may seem to comprise things in some kind of hierarchy of levels – atoms, molecules, cells, organs, organisms, populations, societies, economies, nations, worlds – but actually everything that happens at all those levels really derives from the interactions at the bottom. If you could calculate the outcome of all the low-level interactions in any system, you could predict its behaviour perfectly and there would be nothing left to explain.

There is never only one explanation of anything. We can always give multiple different explanations of a phenomenon — certainly for anything at the macroscopic level — and lots of different explanations can be true at the same time, so long as they are all mutually consistent. Thus one explanation of a leopard’s stalking behaviour will be in terms of the firings of neurons and electrical signals sent to muscles. An equally true explanation would be that the leopard is hungry.

Reductionism does indeed say that you could (in principle) reproduce the behaviour from a molecular-level calculation, and that would be one explanation. But there would also be other equally true explanations. Nothing in reductionism says that the other explanations don’t exist or are invalid or unimportant. We look for explanations because they are useful in that they enable us to understand a system, and as a practical matter the explanation that the leopard is hungry could well be the most useful. The molecular-level explanation of “stalking” is actually pretty useless, first because it can’t be done in practice, and second because it would be so voluminous and unwieldy that no-one could assimilate or understand it.

KM: So, this is what I call the ecumenical version of reductionism which comes in weaker or stronger forms. CH states a weak form above – where he says that explanations at various levels are equally valid. The physicist Sean Carroll espouses a similar view, but states it in what is a subtly stronger (and kind of patronising) way – he describes explanations at higher levels as useful ways of talking about complicated systems – in effect, as convenient fictions. But, he seems to insist that the real explanation is at the lowest level and a description at that level would be the most comprehensive and would necessarily entail and explain all the higher-level organization and apparent causality. (CH makes a somewhat similar argument below).

As a comparison, chess-playing AI bots are now vastly better than the best humans and can make moves that grandmasters struggle to understand. But no amount of listing of low-level computer code would “explain” why sacrificing a rook for a pawn was strategically sound — even given that, you’d still have all the explanation and understanding left to achieve.

So reductionism does not do away with high-level analysis. But — crucially — it does insist that high-level explanations need to be consistent with and compatible with explanations at one level lower, and that is why the concept is central to science.

Mitchell continues:

In a deterministic system, whatever its current organisation (or “initial conditions” at time t) you solve Newton’s equations or the Schrodinger equation or compute the wave function or whatever physicists do (which is in fact what the system is doing) and that gives the next state of the system. There’s no why involved. It doesn’t matter what any of the states mean or why they are that way – in fact, there can never be a why because the functionality of the system’s behaviour can never have any influence on anything.

I don’t see why that follows. Again, understanding, and explanations and “why?” questions can apply just as much to a fully reductionist and deterministic system. Let’s suppose that our chess-playing AI bot is fully reductionist and deterministic. Indeed they generally are, since we build computers and other devices sufficiently macroscopically that they average over quantum indeterminacy. That’s because determinism helps the purpose: we want the machine to make moves based on an evaluation of the position and the rules of chess, not to make random moves based on quantum dice throwing.

KM: There is a strong position in physics and philosophy (espoused forcefully by Bertrand Russell, for example) that argues that causes in fact do not exist. If the universe is really deterministic, then there is no room for and no need for causes – understanding and explanations and “why” questions absolutely would NOT apply. The universe would simply evolve based on the initial conditions and the fundamental forces determining the movements and interactions of all the particles. Determinism thus implies reductionism. Something can only be considered a cause of something else if it being different would have caused a difference to that something else occurring. If there is no way that anything could actually be different from how it is because the universe is evolving deterministically from the dawn of time to the end of time (or indeed because it all just exists in a block universe without a direction of time), then the concept of causation simply does not apply. It relies on counterfactual possibilities being ontologically real. (And so does the idea that information can have causal power in a system).

But, in reply to “why did the (deterministic) machine sacrifice a rook for a pawn” we can still answer “in order to clear space to enable the queen to invade”. Yes, you can also give other explanations, in terms of low-level machine code and a long string of 011001100 computer bits, if you really want to, but nothing has invalidated the high-level answer. The high-level analysis, the why? question, and the explanation in terms of clearing space for the queen, all still make entire sense.

KM: Now, here, as with many compatibilist arguments or thought experiments, I think we need to back waaaaaay up. The high-level explanation only applies in this thought experiment because the computer was programmed by a human being to do things for a reason, and is playing a game developed by humans that has goals and functionalities of components embedded in it. There’s only an answer to the “why?” question because it was programmed in that way. If you are going to posit determinism and try and deduce what follows from it, then you don’t get to just assume the existence of games and computers and computer programmers and go from there. You have to first explain how, in a deterministic universe, you would ever get games and computers and computer programmers.

I would go even further and say you can never get a system that does things under strict determinism. (Things would happen in it or to it or near it, but you wouldn’t identify the system itself as the cause of any of those things).

Mitchell’s thesis is that you only have “causes” or an entity “doing” something if there is indeterminism involved. I don’t see why that makes any difference. Suppose we built our chess-playing machine to be sensitive to quantum indeterminacy, so that there was added randomness in its moves. The answer to “why did it sacrifice a rook for a pawn?” could then be “because of a chance quantum fluctuation”. Which would be a good answer, but Mitchell is suggesting that only un-caused causes actually qualify as “causes”. I don’t see why this is so. The deterministic AI bot is still the “cause” of the move it computes, even if it itself is entirely the product of prior causation, and back along a deterministic chain. As with explanations, there is generally more than one “cause”.

Nothing about either determinism or reductionism has invalidated the statements that the chess-playing device “chose” (computed) a move, causing that move to be played, and that the reason for sacrificing the rook was to create space for the queen. All of this holds in a deterministic world.

KM: Again, I question the validity of the premises of this thought experiment. You have to explain how reasons would ever come to be (how creatures that could have reasons would ever evolve) in a deterministic universe or one where the only kind of causation at play is at the lowest levels of physical forces (which as we’ve seen does not really fit the concept of causation at all). Nothing would be for anything – you would have no purpose, no function, no value, no goals, no meaning.

Mitchell pushes further the argument that indeterminism negates reductionism:

For that averaging out to happen [so that indeterminism is averaged over] it means that the low-level details of every particle in a system are not all-important – what is important is the average of all their states. That describes an inherently statistical mechanism. It is, of course, the basis of the laws of thermodynamics and explains the statistical basis of macroscopic properties, like temperature. But its use here implies something deeper. It’s not just a convenient mechanism that we can use – it implies that that’s what the system is doing, from one level to the next. Once you admit that, you’ve left Flatland. You’re allowing, first, that levels of reality exist.

I agree entirely, though I don’t see that as a refutation of reductionism. At least, it doesn’t refute forms of reductionism that anyone holds or defends. Reductionism is a thesis about how levels of reality mesh together, not an assertion that all science, all explanations, should be about the lowest levels of description, and only about the lowest levels.

KM: As I said above, the version of reductionism I am thinking of (which despite CH’s assertion, many people do indeed assert and seem to hold) is precisely the one that believes ultimately “that all science, all explanations, should be about the lowest levels of description, and only about the lowest levels”. More precisely and more fairly, there are many people who assert the idea that in principle all the important business is happening at the lowest levels, while allowing that in practice it is not possible to fully describe complicated things at that level and so it is much more convenient to work with measurable high-level coarse-grained parameters.

I said above that determinism implies reductionism (no high-level causes, no causes at all, in fact, just the wave function inexorably evolving). But I also claimed the converse, that indeterminacy negates reductionism. This doesn’t obviously follow by necessity, so let me explain the context of that claim. What I was trying to point out in the original post was an inconsistency in the logic of people who allow that quantum indeterminacy exists but claim that it would not percolate up to affect classical levels. At the same time, they maintain a reductionist stance towards explaining the behavior of complex systems and argue for determinism at the classical level. If you admit that coarse-graining happens, from one level to the next, and that not all of the details at the lowest level matter and that many of those details have no effect on the system, then you have just rejected reductionism. You’re not just thinking of the system as a system for convenience, you’re saying that its organization is a key causal factor in its evolution, because this will determine which details matter and which don’t. This is perfectly reasonable and correct, in my view, but it’s no longer reductionist, at least not in a fully orthodox sense. If that can happen from the quantum level to the next one up, then why couldn’t it happen at every transition between levels where some coarse-graining occurs? That would, in my view, be precisely what defines levels and grants them some ontological validity.

So, there’s an “eating your cake, and having it too” quality to that move – it basically invokes macroscopic causation to reject an important role for quantum indeterminacy, while otherwise claiming that all the causal work in complex systems occurs at the lowest (classical) level and rejecting the idea that the organization of the system embodies crucial causal relationships and criteria.

Indeterminism does mean that we could not fully compute the exact future high-level state of a system from the prior, low-level state. But then, under indeterminism, we also could not always predict the exact future high-level state from the prior high-level state. So, “reductionism” would not be breaking down: it would still be the case that a low-level explanation has to mesh fully and consistently with a high-level explanation. If indeterminacy were causing the high-level behaviour to diverge, it would have to feature in both the low-level and high-level explanations.

Mitchell then makes a stronger claim:

The macroscopic state as a whole does depend on some particular microstate, of course, but there may be a set of such microstates that corresponds to the same macrostate. And a different set of microstates that corresponds to a different macrostate. If the evolution of the system depends on those coarse-grained macrostates (rather than on the precise details at the lower level), then this raises something truly interesting – the idea that information can have causal power in a hierarchical system …

But there cannot be a difference in the macrostate without a difference in the microstate. Thus there cannot be indeterminism that depends on the macrostate but not on the microstate. At least, we have no evidence that that form of indeterminism actually exists. If it did, that would indeed defeat reductionism and would be a radical change to how we think the world works.

It would be a form of indeterminism under which, if we knew everything about the microstate (but not the macrostate) then we would have less ability to predict time t + 1 than if we knew the macrostate (but not the microstate). But how could that be? How could we not know the macrostate? The idea that we could know the exact microstate at time t but not be able to compute (even in principle) the macrostate at the same time t (so before any non-deterministic events could have happened) would indeed defeat reductionism, but is surely a radical departure from how we think the world works, and is not supported by any evidence.

But Mitchell does indeed suggest this:

The low level details alone are not sufficient to predict the next state of the system. Because of random events, many next states are possible. What determines the next state (in the types of complex, hierarchical systems we’re interested in) is what macrostate the particular microstate corresponds to. The system does not just evolve from its current state by solving classical or quantum equations over all its constituent particles. It evolves based on whether the current arrangement of those particles corresponds to macrostate A or macrostate B.

But this seems to conflate two ideas:

1) In-principle computing/reproducing the state at time t + 1 from the state at time t (determinism).

2) In-principle computing/reproducing the macrostate at time t from the microstate at time t (reductionism).

Mitchell’s suggestion is that we cannot compute: {microstate at time t } ⇒ {macrostate at time t + 1 }, but can compute: {macrostate at time t } ⇒ {macrostate at time t + 1 }. (The latter follows from: “What determines the next state … is [the] macrostate …”.)

And that can (surely?) only be the case if one cannot compute: {microstate at time t } ⇒ {macrostate at time t }, and if we are denying that then we’re denying reductionism as an input to the argument, not as a consequence of indeterminism.

KM: Okay, so this is the crux of the argument and I really welcome the questioning, which will help me hopefully clarify what I mean. It is not that knowing any particular microstate at time t would not also tell you the macrostate at that time – it certainly would. It’s more a question of understanding the full picture of causality in the system. The reason that any given microstate will tend to lead to some subsequent microstate is by virtue of the macrostate that it entails meaning something.

The existence of indeterminacy makes it possible – not inevitable, by any means, but possible – that systems could evolve (which we call living organisms) that do have goals, purpose, value, meaning, and function – that do things for reasons. Those parameters are encoded at the level of macrostates. Of course, at any given moment, they are realized in some particular microstate, but the low-level details may be incidental. Indeed, many aspects of the microstate of any given neuron or brain region are lost or actively filtered out in the coarse-graining that happens through synaptic transmission and population-level neural dynamics. It’s meaning that drives the mechanism.

Mitchell draws the conclusion:

I agree, but consider that to be a consequence of indeterminism, not a rejection of reductionism.

This brings into existence something necessary (but not by itself sufficient) for things like agency and free will: possibilities.

As someone who takes a compatibilist account of “agency” and “free will” I am likely to disagree with attempts to rescue “stronger” versions of those concepts. But that is perhaps a topic for a later post.

KM: To be clear, the arguments laid out above are only making that case that indeterminacy is necessary for agency to exist (because it creates real possibilities for agents to choose between, and because it opens the door to macroscopic causation). How organisms have evolved to take advantage of that causal slack – to become causes in their own right – is a much longer story, and the subject of my next book ;-) With thanks again to Coel for the discussion.

For more on that argument, see: Mitchell KJ. Does Neuroscience Leave Room for Free Will? Trends Neurosci. 2018 Sep;41(9):573-576. doi: 10.1016/j.tins.2018.05.008.

Does freedom bubble up from the quantum realm?

noreply@blogger.com (Kevin Mitchell) — Sat, 12 Dec 2020 18:36:00 +0000

I’ve been writing lately (in this article and blogpost, for example) about agency and more specifically about whether neuroscience or basic physics rule out the idea of free will, of organisms like us being able to genuinely make choices based on their own reasons and thus act as causal agents in the world. I’m grateful to Philip Goff for responding to some of these ideas in a recent blogpost and to Philip Ball for continuing the conversation in a post of his own.

I respond to their arguments here and make some more general points along the way. (For convenience, I will refer to the two Philips by their last names, which feels a bit rude, so, sorry, Philips!).

The question that Goff takes up is my claim that fundamental indeterminacy at the quantum level creates some room in which free will can operate. He quotes the following passage, which sums up the argument:

“The inherent indeterminacy of physical systems means that any given arrangement of atoms in your brain right at this second, will not lead, inevitably, to only one possible specific subsequent state of the brain. Instead, multiple future states are possible, meaning multiple future actions are possible. The outcome is not determined merely by the positions of all the atoms, their lower-order properties of energy, charge, mass, and momentum, and the fundamental forces of physics. What then does determine the next state? What settles the matter?”

To be clear, this line of thinking does not originate with me. It’s been well explicated by people like George Ellis and Helen Steward in recent times, but really dates back to the ancient Greeks, most notably Epicurus (some of whose work was effectively reproduced by the Roman philosopher Lucretius, in the form of his didactic poem De Rerum Natura (On the Nature of Things)).

One of the most important arguments that Epicurus makes invokes what is now known as “the swerve”. Epicurus inherited from Democritus and his followers the remarkably prescient idea that all matter was composed of simple particles (which Democritus had called “atoms”) and that complicated stuff was made of different combinations of such atoms. But he objected to the rigidly deterministic laws that Democritus had developed to explain how atoms move and what happens when they collide. He famously argued that if there were no element of randomness in the movements of atoms – if they did not occasionally “swerve” – then, first of all, the universe would not exist (which feels like a biggie), and, secondly, that humans, or any animal, would be incapable of autonomous action or real choice.

In modern parlance, if we are made of atoms and their movements are entirely determined by the low-level laws of physics, then the atoms are gonna do what the atoms are gonna do. It doesn’t matter what we want – indeed, wanting would never arise – why would it? Alternate possibilities would simply not exist. Everything would be determined in an unbroken line from the beginning of the universe to the end of time.

The swerve – that element of randomness – breaks that chain and introduces freedom and indeterminacy into the universe. It creates possibilities where none would exist without it. But Epicurus’ target was not just determinism – he also argued forcefully against reductionism, the idea that the origins of causal power, even in complex systems like living organisms, are ultimately entirely traceable to the forces controlling the movements of atoms. He recognised that this idea is incompatible with the autonomy of living beings, including humans.

His argument (and mine) is thus not simply that the world has some randomness in it. Many people rightly point out that randomness by itself does not provide free will. The argument goes that if my decisions are fully pre-determined by the laws of physics, then they are not freely made by me. But if they are determined, in the act of making a decision, by random swerves of atoms in my brain, then they are also not freely made by me.

This is true, but it ignores Epicurus’ wider point – that the existence of some randomness at the lowest levels means that causality does not fully inhere at those levels. Instead, the higher-order configuration of a system can play a causal role in its evolution from state to state. (In other words, when determinism falls, reductionism falls with it).

This is the argument I have been making too, updated with reference to our modern understanding of fundamental physics (and neuroscience, genetics, and evolution). Epicurus’ swerve has a modern echo in the indeterminacy inherent in quantum theory and empirically observed in events at quantum levels. (Though there are multiple different interpretations of what the quantum theory means for the ultimate nature of reality).

The question that Goff and Ball take up is whether this quantum indeterminacy really leaves room for or rescues free will. (As an aside, I don’t think free will needs “rescuing” – the fact that we make decisions is one of the most basic aspects of our experience. It’s what we do all day, every day – go around making one choice after another. Any claim that this deepest aspect of the phenomenology of our existence is an illusion should in my view carry a heavy burden of proof).

Goff takes this indeterminacy (and my position of incompatibilism) for granted for argument’s sake, but says:

“It sounds intuitive, but I don’t think this strategy ultimately works. Even among indeterministic interpretations of quantum mechanics, although the physics doesn’t conclusively settle what will happen, it does determine the objective probability of what will result from any given physical circumstance. Although we can’t predict with certainty, say, where a given particle will be located when we make a measurement, the Born rule tells us, for any given location in the universe, precisely how likely we are to find the particle in that location. It’s not determinism, but it’s not a ‘free for all.’”

He goes on to imagine a scenario where the Born rule plays a role in determining the “objective probability” of his own macroscopic actions:

“Mitchell worries that if the physics determines what I’m going to do, then I’m not really free. But physics determining the objective probability of what I will do is no less constraining. If whether I water Susan (my Madagascan dragon tree) is really up to me – in the strong incompatibilist sense – then surely the physics can’t fix how likely it is that I will water Susan. If it’s just totally up to me, then it could go either way depending on my radically free choice.

Here’s a little thought experiment to make the point clear. Take the moment when I’m about to decide whether or not to water Susan. Let’s say the Born rule determines that there’s a 90% chance my particles will be located in the way they would be if I watered Susan and a 10% chance there’ll be located in the way that corresponds to not watering Susan (obviously this is a ludicrously over-simplistic example, but it serves to make the point). Now imagine someone duplicated me a million times and waited to see what those million physical duplicates would decide to do. The physics tells us that approximately 900,000 of the duplicates will water Susan and approximately 100,000 of them will not. If we ran the experiment many times, each time creating a million more duplicates and waiting for them to decide, the physics tells us we would get roughly the same frequencies each time. But if what happens is totally up to each duplicate – in the radical incompatibilist sense – then there ought to be no such predictable frequency. The number that do and don’t water the plant should change each time, as the radically free choices of each individual varies.”

My first reaction to this is similar to the main point that Ball makes in his own commentary – namely, that it is an ill-posed thought experiment. The Born rule applies to quantum events and elementary particles, not to macroscopic objects and certainly not to the decision-making of agents. Moreover, it applies independently to single particles and events (unless they are entangled). So, in a complex system, all those probabilities will multiply exponentially to create a massive web of indeterminacy.

This is the main point of my argument. The quantum events do not decide the outcome – they merely make the whole system indeterminate. This has two key implications: the next state of the system is NOT determined by the current state plus the laws of physics. And the causal influences are NOT restricted to the lowest level of the system.

I also do not see what the idea of “objective” probabilities of various outcomes being determined by physics is doing in this argument. The whole idea of probability is open to many different interpretations and it’s not clear at all that there is, or should be, a straightforward mapping of its meaning at the level of quantum events to its meaning at the level of human choices. These levels are simply incommensurate. As Ball says, “It’s best, I think, to explain phenomena at the conceptual/theoretical level appropriate to it.”

The important point is that possibilities exist – they don’t have to be equally likely. Indeed, it will almost always be the case that you won’t weight the choices open to you equally. I don’t think anything in particular follows from that.

Goff follows up with a broader point, which is interesting, though I’m struggling to fully grasp what he means by it:

“All is not lost, however. I think Mitchell is conflating two claims:

The laws of physics are deterministic
The universe as a whole is deterministic

How could these come apart? They come apart if the laws of physics are ceteris paribus laws, i.e. laws that tell us what will happen in the absence of other causal influences. On this interpretation of physical law, the probabilities yielded by the Born rule are the objective probability of what will occur in the absence of some other causal influence. Such other causal influences might include the kind of irreducible causal powers Mitchell believes reside at the level of neurobiology. Mitchell seems to be concerned to avoid a violation of the laws of physics. But if the laws of physics are ceteris paribus laws, then higher-level causal powers should be thought of as complimenting the laws of physics rather than contradicting them. If physics basically tells us ‘X will happen unless there are some higher-level causal forces,’ and X doesn’t happen precisely because there are higher-level causal forces, then nothing occurs that is inconsistent with physics.”

I think I kind of agree with this, though I might not word it like that. I do think there are other kinds of causal forces at work, which constrain and inform how the physical state of our brains evolves from moment to moment. In particular, I have argued (in this recent talk, for example, or this longer one) that what drives the next state is the meaning of the current state (or the meanings, plural, of all the sub-states across the brain), which is instantiated in the configuration of the neural circuitry and the history that it embodies.

But I would argue that such an effect or interpretation is only possible because the laws of physics are not deterministic. If they were, then the universe would be deterministic and they would not be ceteris paribus (all other things being equal) laws. There would be, indeed could be, no “other things”, from a causal perspective. (At least, that’s my strong intuition, but I’d be very interested to hear what others, especially physicists or philosophers make of the idea).

Thankfully, it seems that indeterminacy does exist (at least according to a majority interpretation of the current knowledge of physics). Epicurus was right – the atoms (or more fundamental particles) do seem to swerve every now and then. And life, agency, and yes, even what we call free will, can evolve in the causal spaces created by those swerves.

A sinister attractor – why males are more likely to be left-handed

noreply@blogger.com (Kevin Mitchell) — Sun, 22 Nov 2020 19:45:00 +0000

It seems an innocent enough question: why are males more frequently left-handed than females? But the answer is far from simple, and it reveals fundamental principles of how our psychological and behavioural traits are encoded in our genomes, how variability in those traits arise, and how development is channelled towards specific outcomes. It turns out that the explanation rests on an underlying difference between males and females that has far-reaching consequences for all kinds of traits, including neurodevelopmental disorders.

Pixabay.com

A recent tweet from Abdel Abdellaoui showed data on rates of left-handedness obtained from the UK Biobank, and asked two questions: why is left-handedness more common in males and why are rates of reported left-handedness increasing over time?

I don’t think the answer to the second question is known but I presume it has to do with the declining practice of forcing left-handers to write right-handed. This once common practice reflects a long history of prejudice against lefties, illustrated by the derivation of the word “sinister”, which in Latin means “left”, as opposed to “dexter” meaning “right” which is the root of the positive words “dexterity” and “dextrous”.

The first question – why is left-handedness more common in males? – is the one I want to explore here, as it opens up some fascinating questions about robustness, variability, and developmental attractors. This is a long known and extremely robust finding. Rates of left-handedness (from this UK sample) are around 11-12% in males and 9-10% in females and these kinds of differences are seen across cultures and across time.

Why would this be? For that matter, why do we even have handedness? Why is there a bias towards right-handedness? And does this sex difference tell us anything about how handedness emerges or about sex differences in neural or behavioral traits more generally?

Handedness is not unique to humans – many animals show a preference for using one hand over another for (ahem) dextrous movements. (Not so much for big movements or lifting heavy things, but for actions requiring fine motor control). This may reflect the wiring and communication demands of such actions. For fine motor sequences that have to be carried out quickly, it may be much more efficient to code them all in a concentrated area of neural tissue. In the brain, this means avoiding having to communicate back and forth between the two hemispheres. The result is that the neural substrates of fine motor sequences tend to get concentrated in one hemisphere. This may be why control of speech, which involves very sophisticated motor programs, tends to be localised to one hemisphere.

What’s less obvious is why there should be a bias to consistently localise such functions to one specific hemisphere versus the other. One possible explanation is that biasing the process consistently towards one side or the other might make the outcome of hand preference generally more robust. Maybe getting rid of the need to “choose” one side or the other at some stage of development removes a possible failure point from the overall process. (That’s admittedly super speculative, on my part). Others have proposed a variety of social explanations for a consistent directional bias.

Patterning the left-right axis

In any case, in order for the two hemispheres to develop different patterns of neural circuitry (or for our internal organs to be patterned and positioned asymmetrically), the developing embryo must have some systems for distinguishing left from right. In particular, it requires a symmetry breaker, based on something intrinsic in the embryo. The details of how this happens differ between species, but the mechanism ultimately seems to go all the way down to the level of the chirality of various protein molecules, often with some function in the cytoskeleton or cilia motility. After this initial symmetry breaking process, a much more conserved set of signaling pathways is involved in mediating the subsequent differentiation of the left and right sides of the developing embryo, ultimately including the brain.

So, there are genetic processes that reliably code for hand preference (presumably by controlling aspects of neural development that differ between the hemispheres) that may be ancient and found in many animals. And there seem to be separate genetic processes that code for the rightward bias, which may be more unique to humans (but for findings from other animals see here and here).

But why is this rightward bias not 100%? There are two possible reasons: one, a certain rate of left-handedness may be evolutionarily selected for. That is, in a game theory sense, it may be beneficial to be left-handed when most other people are right-handed (in hand-to-hand combat, for example), but once the frequency of left-handedness starts to increase, that advantage decreases.

The alternative is much more prosaic: a certain rate of left-handedness may simply reflect the fact that any genetically encoded process will show some variation, unless that variation literally kills you. (My money is on this explanation because it doesn’t require any additional special reasons).

The genetics of handedness

Whether it’s actively selected for, or just not actively selected against, if genetic variation affects the trait, this means left-handedness should run in the family, and indeed it does. The general rate of left-handedness is about 10%, but if one of your parents is left-handed, then your chance of being left-handed is about 15%. If both of your parents are left-handed, that rate rises to over 20%.

Very large-scale twin and family studies have estimated the heritability of handedness at about 25%. That means, across the population, about a quarter of the variance in handedness is due to genetic differences between people. Recent genome-wide association studies have had some success in identifying genetic variants that predispose (very weakly) to left-handedness. Some enrichment of genes encoding cytoskeletal proteins seems congruent with the known cellular mechanisms of left-right patterning (but frankly it’s early days with these studies).

The important thing here is that the inheritance of left-handedness is probabilistic. In fact, you don’t inherit left-handedness – you inherit a certain likelihood of being left-handed. We can see that this probability plays out independently, even in monozygotic twins. If one twin is left-handed, the chance of the other being left-handed is only about 25% (a good bit higher than the population average but obviously much less than 100%).

So what other factors determine how this genetic likelihood plays out? Could our upbringing have an effect? Well, most parents will probably recognise that the handedness of their children is not something they had any influence over. And the twin and family studies bear this common experience out – they have consistently found no effect at all of the family environment on handedness. Nor have any systematic factors in the wider environment been reliably associated with handedness.

If it’s not genes and not environment, what else could be affecting this trait? Well, there is a third source of variation that is extremely important, though often overlooked, which is development itself. As I have written about extensively (including in my book, “Innate”), the processes of development are noisy, in engineering terms. What is encoded in the genome is not a highly specified outcome but rather a set of biochemical rules that underpin the processes of development. When played out, these mindless biochemical algorithms will tend (quite robustly) to produce an outcome within a well-functioning range.

Natural selection frankly doesn’t care that much about all the details – the thing just has to work. The consequence of the inherent noisiness of developmental processes is that every time you run the developmental program from any particular genome, you get a unique outcome. For many traits (like height, for example, or facial morphology), this developmental variation manifests quantitatively. For traits like handedness, however, which is effectively dichotomous, it manifests probabilistically.

A rolling stone

A useful way of thinking about this is provided by the visual metaphor of the “epigenetic landscape”, introduced by Conrad Waddington in 1957. The idea is that a developing organism (the ball) is channelled along certain phenotypic trajectories as it proceeds through development. The shape of the landscape and the relative depth of these channels is determined by the genetics of the individual.

For a landscape derived from an “average” genome, it should be less likely for the ball to roll into the channel that leads towards left-handedness than the one towards right-handedness. If you run 100 balls over this landscape, on average only 10 of them would end up in the lefty channel, while 90 would end up in the righty channel. But if you generated the landscape from the genome of someone who is actually left-handed, the channel towards left-handedness would be easier to get into, and the outcome of 100 balls might be 25:75.

Whatever the shape of the underlying landscape, the actual outcome in any individual would depend on a significant element of chance. At some point in development, represented by the arrows in the diagram, a small amount of noise might nudge the ball towards one or other trajectory. The self-organising processes of development would then reinforce that choice to ultimately produce dichotomous outcomes. In the parlance of dynamical systems, left- and right-handedness are two distinct phenotypic attractors.

The development of handedness may rely on processes that act like a switch. We saw above that it is important to localise fine motor control to one side of the brain or the other. It may not matter that much whether it’s left or right – what you want to avoid is failure to localise it at all. Left- or right-handedness may thus represent mutually exclusive – even possibly antagonistic – developmental trajectories. “Choosing” one pathway may inhibit the other and reinforce itself. This view is consistent with the fact that left-handedness is much more common (and thus presumably much better tolerated) than either mixed handedness or non-handedness.

What this all means is that, though handedness is only partly heritable, it may be largely or even completely innate. (As an aside, a similar situation may hold for other traits, such as sexual preference).

Now, back to the question we started with – why should sex affect the outcome of development with respect to handedness? What is about the way males develop that would lead them to end up left-handed more frequently, and why should things be this way? There are many scenarios one could come up with that invoke some sort of special connection between sex and handedness, maybe involving sexual selection, males with rare phenotypes having an advantage, and so on. These all seem very speculative to me and I don’t think we need to posit any such mechanisms. A more likely explanation to me, and one that is certainly more parsimonious, is that the increased rate of left-handedness reflects the greater male variability that is observed in many traits, not just in humans, but in many other animals too.

Greater male variability

Males show a small but statistically significantly higher variance than females for all kinds of traits in humans. The spread of values is greater in males for birth weight, morphological traits, a range of blood parameters, and even things like athletic and academic performance. The same is true for IQ, with males showing a wider, flatter curve than females, with a consequently greater proportion of males at both the low and high ends. There are similar findings in mice, especially for morphological traits, though females showed a greater variance in immunological measures, possibly due to fluctuating hormonal influences.

You can see this kind of variability play out within an individual too, across the two sides of the body (which represent largely independent “runs” of the same genomic program). Males show higher levels of “fluctuating” (i.e., random) asymmetry of facial morphology than females.

Why would this be? Well, first, the fact that this trend is observed across species argues strongly against a cultural origin for it, though initial biological differences might well be amplified through cultural mechanisms in humans. But what would cause it in the first place? We might look to hormonal influences – maybe testosterone just plays havoc with the precision of developmental programs, for example. But as we will see below, the same phenomenon is observed in insects where sex hormones play no role in morphological differentiation.

Something more fundamental is going on and it can be traced to the sex chromosomes – in particular, to the fact that males in mammals have only one copy of the X chromosome, while females have two. Genetic variation on the X chromosome will therefore have a bigger phenotypic effect in males, not just for single loci, but as a collective effect. This paper by Reinhold and Engqvist sums it up nicely:

“In females, the traits that are influenced by X chromosomal genes will be under the average influence of the two parental copies, whereas in males, the effect of the single X-chromosome will not be averaged. As a result, male mammals are expected to show larger variability than females in all traits that are, at least to some extent, influenced by X-chromosomal alleles.”

This is not to suggest that all the traits mentioned above are X-linked, in the normal sense of that term. In fact, they are all highly polygenic, involving genes all over the genome. It is just that some of those genes are on the X chromosome and variation in such genes will be averaged out or buffered more in females than in males. Because the buffering of genetic variation is a little less efficient or robust in males, the variance of the traits will be consequently greater.

Now, the hypothesis that this effect is due to the number of X chromosomes, rather than effects of male hormones, makes a strong prediction, which Reinhold and Engqvist tested beautifully, taking advantage of the fact that in some species females are the ones with two different sex chromosomes. This is the case in many bird species, where males have two Z chromosomes and females have a Z and a W. A similar situation holds in butterflies. But most mammals and many insects have an XX (female) and XY (male) arrangement.

The prediction is that in such species, the effect should be reversed – females should be more variable. Looking across many different species of mammals, birds, butterflies, and insects, a pattern consistent with the sex-chromosome hypothesis was found: the sex with two different sex chromosomes was more variable. (As an aside, this paper is a wonderful example of how to do science – an insightful and bold prediction, allowing a stern test of a hypothesis, with painstaking research to follow it up).

Okay, back to humans. One way to interpret the observation that males are more variable is that they are less robustly channelled into narrow phenotypic outcomes. As discussed above, this will mostly manifest as quantitatively greater variance for all kinds of traits. For handedness, it should mean that developing male organisms are less strongly channelled into the trajectory leading to right-handedness. In our visual metaphor, this may mean a slightly lower ridge between the channels for a longer time or generally a flatter plain preceding the important choice point. Development in males should be just a little noisier and all the feedback loops that normally ensure robust development should be just a little less able to buffer that noise. In effect, male embryos may be slightly more likely to jiggle into the left-handed channel.

[Note: I’d love to try and formalise those intuitions or even have a virtual marble run to play with parameters in, if there are any modellers out there who’d fancy having a go].

If this were just about handedness, it’d be a cute story, but maybe just a curiosity. But we can see the same effect on other phenotypes, including susceptibility to neurodevelopmental disorders. Males have substantially higher rates of diagnosis of a range of such disorders, including autism, ADHD, schizophrenia, stuttering, tic disorders, and others. These conditions can be caused by rare mutations in any of a very large number of different genes, with complicated polygenic background effects at play.

A now well-replicated finding is that females with such diagnoses have more serious mutations on average than males do. This suggests that it takes a bigger mutation to push a developing female brain into one of these phenotypic outcomes, while developing male brains are less able to buffer the effects and therefore more vulnerable. Consistent with this view, when such mutations are inherited from a parent (as opposed to arising de novo in the generation of sperm or eggs), they have been found to be more likely to have been inherited from the mother. One interpretation of that result is that males who had such a mutation would have been more severely affected and thus less likely to have children, while females are better able to buffer the effects.

See what happens when you ask a seemingly innocent question? ;-)

Corvid consciousness – computation, cognition, or comprehension?

noreply@blogger.com (Kevin Mitchell) — Wed, 30 Sep 2020 08:03:00 +0000

A really nice paper came out recently that claims to

have discovered a neural correlate of sensory consciousness in a corvid bird (the carrion crow). The authors use an elegant set up involving barely perceptible visual stimuli to distinguish the delivery of a stimulus and the subjective percept that it engenders. The experiment clearly demonstrates that crows can maintain an internal representation for a period of time before taking an action based on a rule that is subsequently presented to them. This kind of task has been used in primates to distinguish what happens in the brain when an animal consciously detects a stimulus versus when it doesn’t. But is this really a correlate of conscious subjective experience or simply a marker of ongoing neural activity that mediates working memory? What do we even mean by conscious subjective experience? Does maintaining an active neural state necessarily entail a mental state?

The experimental set up is really powerful. The authors trained two crows to perform this behavioral task while recording from hundreds of neurons in a particular part of the bird brain that is thought to be analogous to the mammalian prefrontal cortex. The birds are either shown a stimulus or not, for a very brief time – so brief that they sometimes detect it and sometimes don’t. Whether they do or not is determined in some way by the current internal state of the crow’s brain (this may reflect the precise phase of rhythmic patterns of ongoing neural activity and excitability at the moment of stimulus presentation).

After a brief period, the crows are then presented with a rule, which tells them how they should act in response to either the presence or absence of a stimulus. This means that the crow must maintain some sort of internal representation for some period of time while waiting to be told what to do with that information.

The important thing about this set up is that it dissociates the presence of the stimulus with the percept of the animal. In particular, the crows can be wrong in two ways – they can fail to see something when a stimulus was actually presented, or they can think they saw something when no stimulus was presented, leading to a distinction between direct visual stimulation and subjective experience.

In primates, including humans, you get quite different brain responses on trials when the subject reports (either by behavior or by language) that they consciously saw something versus when they didn’t. Neurons in primary visual cortex respond to a visual stimulus whether the animal “sees” it or not. Trials where the animal reports a percept involve subsequent activation of many other areas of the cortex, especially frontal regions. This suggests that for a visual stimulus to rise to the level of conscious awareness, it must be broadcast to the rest of the brain.

The experiment with the crows was looking for the same kind of marker or neural correlate of conscious perception. And indeed they found it. Neurons in the posterior pallium of the crows (the nidopallium caudolaterale, for those keeping score at home) responded in two ways: some neurons showed an immediate response to the visual stimulus whether it was perceived or not. Others showed prolonged activity after the stimulus was removed, which correlated much better with the subsequent behavior of the animal than with the stimulus itself.

Importantly, because the crows don’t know what they’re supposed to do with the information during this waiting period, this neural activity should not be interpreted as merely preparation for a particular motor action. Instead, the authors interpret it as reflecting a subjective conscious signal that is maintained over this period. This interpretation is supported by the fact that the patterns match the animal’s percept (as reported by its behavior), even when it was mistaken.

So, what does all this tell us? The similarity with the situation observed in primates prompts the authors to conclude that crows are similarly having some kind of conscious experience. At least, they take these data as evidence for “access consciousness”, meaning that some internal neural state representing something in the outside world can be maintained over time even in the absence of that stimulus and that this information can be made available or broadcast to the rest of the brain and subsequently acted upon.

They are more cautious about taking it as evidence for “phenomenal consciousness”, which would entail it feeling like something to have that experience. Maybe it does, maybe it doesn’t – it’s not obvious at all that it has to. And of course for animals without language, it’s almost impossible to answer that question – if they can’t tell you that it feels like something, it’s hard to find some other way to tell whether it does or not.

The question is whether it’s justified calling this consciousness at all. Maybe it’s just working memory – maybe some neural circuits can maintain reverberating patterns of activity over certain periods of time, can communicate that activity to other parts of the brain, and can use it to inform action. Does any of that have to entail subjective experience? I would guess that lots of animals can probably do that kind of task, possibly even very simple ones (though readers may enlighten me otherwise). Indeed, it seems like you could design circuits that would work like that in a robot without expecting it to subjectively perceive anything. Could this all be achieved by mindless computation?

Or maybe the alternative is true – maybe that kind of internally sustained representation defines mental experience. Maybe you don’t need to add anything else to that – maybe having such representations and being able to act upon them in response to further information merits being called cognition. But does that necessarily entail subjective experience? We can have cognition without consciousness so is there anything else here that justifies the grand conclusion from this paper: that crows are conscious?

Well first there’s the question of what is being represented. You might think it’s the visual stimulus but it isn’t just that (it may not be that at all, after the initial brief moment) – it’s the belief of whether such a stimulus was present or not. (Remember, the crow can be mistaken). The crow knows that it will receive instructions to behave in one way or another depending on whether it perceived the stimulus. It doesn’t have to maintain a trace of the visual stimulus itself –which happens to be a grey square – just a record of its own perceptual decision.

As an aside, many people would quibble with the word “representation” here, and for good reason. That term is used in so many different and often only vaguely defined ways across neuroscience, cognitive science, psychology, computer science, and philosophy. Here I think it is justified, as it refers to a sustained pattern of activity that reflects a belief about the world and that can be communicated or “re-presented” to other parts of the brain to inform action.

This seems to go beyond a simple kind of working memory in that what is being remembered or maintained is an inference, not merely a stimulus. Does this thereby qualify as mental activity? Would it qualify as comprehension? Does the crow know what it knows? Or is it just higher-order neural activity? Again, it seems like you could program that kind of perceptual decision-making and retention of an abstracted signal of the decision for some time in a robotic system, without entailing mentality.

The other characteristic relevant to questions of consciousness is that this representation is made available to the rest of the brain. This is reminiscent of the Global Neuronal Workspace theory of conscious awareness developed by Stan Dehaene and colleagues. The idea is that perceptual stimuli have to reach some threshold of neural activity in order to ignite activity across a much wider network of the brain. In such cases the stimulus is actually perceived consciously. When this ignition fails to occur the stimulus is not perceived, though we can see that lower areas of visual cortex respond the same in either case.

(There is, incidentally, congruent work on conditions like face blindness and tune deafness, which suggests that the problem in these conditions is not the primary responsiveness of face or music processing areas of the cortex, but rather the communication of signals from these areas to frontal regions, necessary for conscious processing of those stimuli).

More broadly, in this hypothesis, percepts are in some

sense competing with each other for access to this global workspace, this unified consciousness. It is in this global workspace that not just percepts but also memories, ideas, beliefs, goals, plans, etc., are integrated, selectively attended to, and acted upon to drive coherent and sophisticated behaviour.

Is that the case in these crows? Who knows? The experimental set up does not provide much scope for sophisticated behaviour – they have been trained to respond in simple ways to this isolated information, not to integrate it with other information and operate upon it in any more complex way. However, the well-known problem-solving abilities of crows (e.g., here) suggest a highly developed capacity for reckoning and figuring of various kinds.

By themselves, the data in this paper can’t really address the question of subjective conscious experience in these birds, limited as they are to this very simple set up. But they’re certainly consistent with some kind of mental life. Perhaps the ability to derive and maintain and broadcast these kinds of perceptual beliefs (or other cognitive features that can be abstracted away from direct current experience) is a necessary but not sufficient function for mental experience.

Maybe that kind of activity would only entail subjective experience in an organism with a broader sense of self, phenomenologically calibrated through its own history of experience with a wider world. It seems most likely that such capabilities exist in graded fashion across the animal kingdom, not all-or-none. It’s probably an ill framed question to ask: are crows conscious? Perhaps better to ask what their consciousness is like.

A neural correlate of sensory consciousness in a corvid bird. Andreas Nieder, Lysann Wagener, Paul Rinnert. Science 25 Sep 2020: Vol. 369, Issue 6511, pp. 1626-1629
DOI: 10.1126/science.abb1447

Are bigger bits of brains better?

noreply@blogger.com (Kevin Mitchell) — Thu, 27 Aug 2020 17:20:00 +0000

We scoff at the folly of phrenology – the simplistic idea that the size and shape of bumps on the skull could tell you something about a person’s character and psychological attributes. It was all the rage in the Victorian era (the early to mid-1800s) in the UK and the US especially, with practitioners armed with calipers claiming to measure all kinds of personal propensities, from Acquisitiveness and Combativeness to Benevolence and Wonder. The skull bumps were just a proxy, of course – the idea was that they reflected the size and shape of the underlying brain regions, which were what was really associated with various traits. It all seems a bit quaint and simplistic now (apart from the entrenched association with racism), but while we may like to think we have moved on, a lot of modern human neuroscience is founded on the same premises.

- The first premise is that different mental functions or psychological traits can be localised to specific regions of the brain.

- The second is that the size of those regions is correlated with the level of function or trait – usually with the idea that bigger is better.

- And there is a third premise, which also sometimes comes along for the ride, which is captured by the slogan “use it or lose it” – the idea that if some brain area or function it supports is not used, that brain area will get smaller; the converse idea is that if it is used more, it will get bigger.

These ideas are pervasive in the public perception of neuroscience, promoted, in my view, by the way we neuroscientists tell our stories. But this is not just a problem of communication of science – these assumptions are also implicit and typically unexamined in the motivation for and interpretation of many studies in the field.

They weren’t plucked out of the air, of course – there is some underlying truth to all of these ideas, and a relevant evidence base supporting them. For example, lesion studies, patterns of brain activation during various tasks, and the selective effects of stimulation of various bits of the brain all support the partial localisation of all kinds of cognitive functions. However, none of these kinds of evidence suggest that the implicated brain regions carry out these functions by themselves. We now understand that most cognitive functions are mediated by distributed networks, not by isolated brain regions. There is certainly a high degree of specialisation of function in that extended circuitry and it’s interesting and important to map that out, but I think we’d all hope to have moved beyond naïve “blobology”.

Similarly, while there are plenty of examples where size really does matter for some function or other, in both animals and humans, it is not obvious that equivalences can be drawn across the range of scenarios where this seems to hold. Differences in size and function (and correlations between them) have been studied in many species across evolution, across development, across the lifespan, across individuals, between sexes, across seasons, in pathological states, and in response to experience. Findings from one of these areas are often used to bolster claims or motivate experiments in another, but disparate underlying mechanisms may be at play across these diverse scenarios.

This is especially true in human studies as the only non-invasive way to measure the size of bits of brains in humans is by neuroimaging, which is extremely crude, relative to techniques that can be applied to animal brains. Methods like voxel-based morphometry (VBM) let you gauge the size of different bits of the brain across individuals by warping 3D images of their brains to a common template. You can similarly measure thickness or surface area of bits of the cortex and compare across individuals. But there are all kinds of different parameters at a cellular level that could drive variation in the overall size of bits of the brain, which may be indistinguishable by neuroimaging.

These parameters include numbers of neurons, numbers of astrocytes and oligodendrocytes, cell sizes, the extent of dendritic or axonal arbors, numbers of synapses, myelination, vascularisation, or many others. And each of those parameters is itself determined by multiple distinct processes, such as proliferation, differentiation, migration, survival, growth, pruning, and on and on. Very different processes and parameters will be at play in different situations.

To take a few examples, a difference in size of some region over evolution may be caused by changes in proliferation and neurogenesis, which could in turn have various underlying causes, such as direct effects on cell cycle genes or indirect effects caused by differences in afferent connectivity and the supply of growth factors; changes with aging may track death of neurons or glial cells, loss of myelin or other degenerative processes; while differences across individuals could be due to variation in neuronal number or in things like the elaboration of neurites and synapses – the same number of neurons, but with bushier connections. All of these may manifest as a difference in size of some region by neuroimaging, but they’re not clearly meaningfully related to each other.

The causes of size differences are thus quite diverse. What about the consequences? Is bigger necessarily better? There certainly are cases where it is, but again, it’s important not to conflate these or compare apples with oranges.

When it comes to the whole brain, size certainly matters. Brain size correlates well with intelligence across species, with bigger-brained species mostly being more intelligent, and especially a consistent trend of increasing size in the primate lineage leading to humans. And if we look across humans, whole brain size also correlates with intelligence (measured in any of a variety of ways), with correlation coefficients in the range of r=0.2–0.3.

There is an interesting exception to that, however. Average male brain size in humans is about 10% larger than average female brain size, but there is no difference in mean I.Q. between the sexes. This suggests some other sex difference (in functional organisation, perhaps) counteracts the size effect on I.Q., which otherwise holds within each sex. Even for the whole brain, size clearly isn’t everything, a principle which holds true across species too.

Whole brain size is also obviously the grossest neuroanatomical measure you can get. What about smaller bits of the brain? Does the size of any of them correlate with function or variation in traits? If you’ve been reading the human neuroimaging literature for the last couple of decades or the popular media describing it you would certainly be convinced that many examples exist of such a relationship.

I would guess there are thousands of published papers reporting such an association. But I can’t think of any that have robustly replicated and there is good evidence to suggest that most such reports are spurious findings.

A very recent paper by Marek and colleagues provides compelling evidence that most such studies have been statistically underpowered by a couple orders of magnitude. Where sample sizes have typically been in the tens or low hundreds, this paper shows that you need to get to samples in the range of 10,000 people before you can detect robust associations between neuroimaging measures and behavioural phenotypes. Below that, sample variance leads to wildly fluctuating results, producing many false positive “findings” and hugely inflated effect sizes. The pernicious effect of publication bias massively exacerbates the problem in the literature. (The similarities to the candidate gene association literature are not coincidental).

Moreover, even for the positive associations that were found in this study, they were with extremely broad psychological phenotypes – general cognitive ability and psychopathology – and the anatomical measures associated with them were highly distributed across the whole brain. What about more specific associations between the sizes of bits of the brain and more specific traits or functions? Is there any reason to think they should exist?

Clearly, we could reasonably expect that the variation that we observe in psychological traits is caused by variation in some parameters in the brain, but should we expect them to be manifested at the macroscopic level probed by neuroimaging? Or are we just taking that approach in humans because no other methodology is available? Are we doing it just because it’s the only thing we can do, not because there is a principled reason to expect such associations to exist?

If the entirety of the prior literature reporting such associations is now in question, is there any other evidence base for the hypothesis that the size of specific brain regions is associated with degree of function? Well, yes, though most of the evidence that springs to my mind is somewhat indirect, for example from pathological situations, from studies looking over evolutionary timescales, or from seasonal changes in some animals that may be quite unusual.

In many neuropathological conditions, or even in normal aging, the degree of loss of grey or white matter often correlates with the severity of neurological or psychological symptoms. However, as with lesion studies, there is a limit to how far this kind of relationship can be extrapolated – greater loss of neural tissue in pathological situations is certainly worse, but it’s not clear that having more in a given area in a healthy situation (i.e., across individuals in the healthy population) would necessarily be better for some specific function.

In evolution, there are lots of examples of differences in size of brain regions that correlate with behavioural adaptations across species. For example, in mammals that have evolved to use various senses to different degrees, those senses that are used more have a larger representation in the cerebral cortex. Different species may literally need more neural real estate to effectively process the richer signals coming in through the visual or auditory or olfactory domains. There’s no point having bigger eyes with more photoreceptors of different kinds enabling more sophisticated visual processing if the brain can’t handle that information.

It’s interesting to note that the underlying mechanisms involve a dynamic communication between different parts of the nervous system, from the sensory periphery through all the relays to the higher perceptual centres. The size of any given brain region is not determined solely by a genetic program that runs in isolation in those cells – it is also affected by proliferation signals and neurotrophic factors released by incoming nerve fibres from other regions in the circuit. Thus, over both development in individuals and the evolution of species, the size of functionally coupled regions is tightly coordinated. (Indeed, cortical thickness within individuals is correlated between regions that make up extended functional circuits).

An even more dynamic relationship between size and function is observed in some animals with seasonal changes in the size of some brain region. For example, in some male songbirds, the high vocal centre or “song nucleus” grows in response to changes in day length and hormonal signals, giving them more neural resources for the mating season, when their singing needs to be on point. Similarly, some mammals and birds show changes in the size of the hippocampus across seasons, which may reflect different behavioural demands on navigational abilities across the seasons. Such changes involve high levels of neural turnover and proliferation of new neurons, tightly regulated by a variety of signaling mechanisms.

These seasonal changes thus represent good examples of all three premises: particular behavioural functions in these species do rely (especially though not exclusively) on these specific brain areas; the size of these areas does seem to reflect degree of functional capability; and when they are not needed at full capacity they shrink (presumably due to an otherwise unnecessarily high metabolic cost of maintaining them).

Though humans don’t show such seasonal changes, there are numerous reports of differences in brain region size that are supposedly driven by differences in experience and associated with the level of some cognitive function. The most famous of these reports involve London taxi drivers, who were found to have larger hippocampi (specifically the posterior portion) than non-taxi-driving subjects. The interpretation was that this difference was driven by their intense training on the geography of London streets (“the knowledge”), with the implication at least that this increase in size improved their navigational abilities.

[Note added: See also a subsequent paper finding a change in hippocampal size following training:

Acquiring “the Knowledge” of London's Layout Drives Structural Brain Changes]

These studies are very well known and effectively taken as lore, both in the field and by the general public. (I have had taxi drivers themselves tell me about these findings). Without dwelling on it, I think the design of the original studies would not stand up well to current expectations, especially in terms of sample sizes, which were in the tens (not the tens of thousands). Moreover, larger studies have not observed a relationship in the general population between size of the hippocampus (or bits of it) and navigational abilities.

Again, the rationale for the hypothesis that bits of the brain should grow with use is rather vague. Are we expecting new neurons to be produced? (This might be the case in the hippocampus though even there the question of whether adult neurogenesis actually occurs in humans is controversial). Or would it be driven by growth of dendrites and synapses? Or maybe very active regions get swole by attracting a bigger supply of blood vessels?

But what does “very active” even mean here? This idea harkens back to the myth that we only use 10% of our brains. Really we’re using all of our brains, all the time, and even when we’re not obviously actively “doing something” with some brain region, there’s still loads of background neural activity there.

Are the hippocampi of taxi drivers really more active than those of other folks? I don’t just mean that they are using them more to navigate – I mean that if you measured the actual neural activity in those structures over some period of time, it would be on average higher (to an extent that might plausibly drive a cellular physiological response and growth of the whole area). That might be true but I don’t think it’s been shown.

Or are they in some way differently active? Maybe there’s more high-frequency firing or more synchronised firing or variation in some other global parameter without a change in overall number of spikes. Perhaps cells could monitor and respond to such parameters. There is an extensive body of literature in animals showing that both levels and patterns of neural activity do feed back onto gene expression in ways that are important for both activity-dependent development and plasticity. As far as I know, however, these have not been linked to macroscopic growth of whole areas but rather to highly cell-type-specific microscopic changes in connectivity.

More generally, it is true that synaptic plasticity and learning, especially the formation of long-term memories, involves the physical growth of new synaptic connections. However, this is not one-way: synaptic plasticity equally involves the weakening and pruning of connections under other conditions. Moreover, absolute increases in overall numbers or strength of synapses are actively counteracted by homeostatic processes that reduce overall synaptic connectivity (especially during sleep), maintaining relative changes in strength, while renormalizing the responsiveness of the network to new experience.

In short, brain is not like muscle. Bits of brain don’t just grow with experience – they mainly change by reorganising their internal connectivity. This is just as well because if the brain did continue to grow with use, all of our brains would be busting out of our skulls. I don’t mean to be too sarcastic (just the right amount), but I’ve been going around seeing things like crazy – really intensely using my visual system for many years now – without causing massive growth of my visual cortex.

So, where does all this leave us? Let’s reconsider our three premises:

Can we localise functions to specific brain areas? We can certainly implicate areas as being involved more in or required more for some functions than others, without falling into the trap of thinking that they are somehow sufficient for those functions. (You can say the same for genes).

Is bigger better? In many scenarios – evolutionary, developmental, pathological – the answer is clearly yes. I’m sure that readers will think of many other examples that fit with this general idea. But to return to where we started, with phrenology, those observations do not necessarily imply that variation in size of bits of brain should contribute to variation across the normal range of behavioural traits or cognitive functions within humans.

Should we expect variation in extraversion or impulsivity or murderousness or spatial reasoning or working memory or creativity be associated with the macroscopic size of specific brain regions, as measurable by neuroimaging? Maybe. It’s not an entirely daft idea, it’s just really crude and simplistic, in my opinion. An alternative hypothesis that I find more likely is that variation in those kinds of complex psychological traits and cognitive functions will be due to idiosyncratic combinations of distributed and likely subtle effects on all kinds of cellular and biochemical parameters affecting the function and computational operations of highly extended circuitry across the brain. Complex traits are called that for a reason.

And finally, do bits of brain grow with use in a way that mediates greater function? The motivation for this hypothesis is vague and weak, in my opinion, and though it is taken as lore, the evidence for this kind of effect in humans is on far shakier ground than many people both in the field and in the general public appreciate.

Generally speaking, while our technology has improved, we won’t make progress in understanding the complex relationships between human brain structure and function and our individual psychological traits armed only with Victorian-era theories.

Escaping Flatland - when determinism falls, it takes reductionism with it

noreply@blogger.com (Kevin Mitchell) — Fri, 31 Jul 2020 10:42:00 +0000

Reductionism is related to determinism, though not in a straightforward way. There are different types of determinism, which are intertwined with reductionism to varying degrees.

The reductive version of determinism claims that everything derives from the lowest level AND those interactions are completely deterministic with no randomness. There are things that seem random, to us, but that is only a statement about our ignorance, not about the events themselves. The randomness in this scenario is epistemological (relating to our knowledge or lack of it), not ontological (a real thing in the world, that we can observe, but that does not depend on us for its existence).

That’s the clockwork universe – the one where Laplace’s Demon (an omniscient being) could unerringly predict the future of the entire universe from a fully detailed snapshot of the state of all the particles in it at any given instant. It’s pretty boring, that kind of universe.

There is also an ostensibly non-reductive flavour of determinism, which simply affirms that every event has some antecedent physical cause(s). Nothing “just happens”. Under this view, however, causes don’t have to be located solely in the interactions of all the particles or limited to the actions of basic physical forces (which also act at the macroscopic scale, determining the orbits of the planets, for example). The causality, or some of it at least, could inhere in the organisation of a system and the constraints that it places on the interactions of its constituents. In this kind of scheme, there is room for a why as well as a how.

That’s the argument, at least, though it’s a little incoherent, in my view. If there is no real randomness in the system (or in the universe as a whole), then I don’t see how you can escape from pure reductionism. In a deterministic system, whatever its current organisation (or “initial conditions” at time t) you solve Newton’s equations or the Schrodinger equation or compute the wave function or whatever physicists do (which is in fact what the system is doing) and that gives the next state of the system. There’s no why involved. It doesn’t matter what any of the states mean or why they are that way – in fact, there can never be a why because the functionality of the system’s behaviour can never have any influence on anything. I would go even further and say you can never get a system that does things under strict determinism. (Things would happen in it or to it or near it, but you wouldn’t identify the system itself as the cause of any of those things).

But what if determinism is false? Let’s see what happens to reductionism when you introduce some randomness, some indeterminacy in the system. Of course, this is exactly what quantum theory does, at least under one interpretation, though there is deep disagreement among physicists as to whether the randomness observed at quantum levels is epistemological or ontological. But let’s say it’s the latter – that randomness really exists in the universe – that some things, at very small scales at least, do “just happen”.

What effect does that have on things at big scales – the scales of rocks and cats and babies, and other things we care about? Some people argue – rather casually, in my view – that randomness at quantum levels will not have any effect at the level of classical physics, because all that noise will be somehow absorbed or averaged out in the system and will not percolate up to higher levels. This means the behaviour of the system at classical levels can still be considered to be deterministic.

Is that true? Do quantum effects stay there at the quantum level? I can think of lots of instances where they wouldn’t – like Schrödinger’s famous cat, for example, whose fate was to be determined by the random decay of a radioactive atom. And I would guess that the randomness of quantum phenomena has some important implications for real-world quantum computing.

But just speaking philosophically, what’s strange about this assertion – as I said, often thrown out very casually – is that it betrays the very idea of reductionism. It relies on the idea that reality is not in fact flat. It claims explicitly that things can be happening at the lowest level of the system that do not percolate up to higher levels. Heresy! How can a good reductionist believe that? Does reductionism only apply at classical levels? Does it stop being true at some scale? Why?

If the properties at the classical level derive from interactions at the quantum level (and we know they do because they can be derived from quantum theory), then why would the subset of such interactions that happen to have arisen randomly not also manifest at higher levels? How would the system know which ones were random and which were determined?

I know that’s a silly way to put it, but it highlights something crucial – the idea that something important is happening at the level of the system. You might say that the reason those quantum fluctuations don’t manifest at the level of the whole system is because they average out. They are random, after all, and if there are many of them and they are independent, then their collective effects should cancel each other out. But that relies on a very non-reductionist mechanism: coarse-graining.

For that averaging out to happen, it means that the low-level details of every particle in a system are not all-important – what is important is the average of all their states. That describes an inherently statistical mechanism. It is, of course, the basis of the laws of thermodynamics and explains the statistical basis of macroscopic properties, like temperature. But its use here implies something deeper. It’s not just a convenient mechanism that we can use – it implies that that’s what the system is doing, from one level to the next.

Once you admit that, you’ve left Flatland. You’re allowing, first, that levels of reality exist. And second, that what happens at one level is only a coarse-grained, statistical reflection of what is happening at the level below.

In my view, that’s almost but not quite right. Any hierarchical system is averaging, or integrating over, or in some way coarse-graining the low-level details, but not just the random ones (how could it be?) – it’s coarse-graining ALL of them. And not just from the lowest, quantum level, to the next one up. This happens at EVERY level. It may be turtles all the way down, but it’s not turtles all the way up.

The macroscopic state as a whole does depend on some particular microstate, of course, but there may be a set of such microstates that corresponds to the same macrostate. And a different set of microstates that corresponds to a different macrostate. If the evolution of the system depends on those coarse-grained macrostates (rather than on the precise details at the lower level), then this raises something truly interesting – the idea that information can have causal power in a hierarchical system, and, more generally, in the universe.

Some criteria embodied in the structure of the system itself drive a different response to these two macrostates. Simple versions could involve a threshold effect, such as a thermostat triggering a heater if the temperature drops below its set point, or a neuron firing an action potential if the voltage across its membrane is high enough. That kind of control is inherently informational. In philosophical terms, it relies on counterfactuals being ontologically real – that is, the current state can only carry causally effective information if in fact it was actually possible that it could have been different.

That little bit of indeterminacy is thus key – otherwise it doesn’t matter what the microstate corresponds to as the system is simply going to follow a deterministic trajectory. But I’ve just been talking about those random events being coarse-grained, along with all the non-random events, so how could they lead to different macrostates? The answer is they’re averaged but not always averaged out. Sometimes those random events will make a crucial difference, especially for a system poised at the boundary between two macrostates. In fact, such a scenario actually amplifies small random fluctuations, by causing a qualitative change in macrostate.

In complex, dynamical systems that are far from equilibrium, some small differences due to random fluctuations may thus indeed percolate up to the macroscopic level, creating multiple trajectories along which the system could evolve. This brings into existence something necessary (but not by itself sufficient) for things like agency and free will: possibilities.

What this means is that causation does not reside simply at the lowest levels and the basic laws of physics, nor is it completely instantaneous. The system will not evolve along a single pre-determined line, nor will its evolution simply follow a random path in a tree of possibilities. Instead, in some types of systems – like living organisms – how the system evolves will depend on what those various macrostates mean. What do they correspond to or reflect in the environment, what consequences are they linked to in terms of action, what feedback does the organism get on the outcomes of those actions, and how does that feedback alter the configuration of the system to set criteria for processing that information in the future?

By building up, not out, creating a functional hierarchy of levels within their own structure, and incorporating meaning in feedback loops that extend through action and consequence into the environment and over time, evolution has created organisms that use the wiggle room provided by stochasticity to exert “top-down” causal power to do things for reasons. The organism itself can choose among those branching possibilities. (More on that here and much more to come).

To come back to where we started, while they are often presented as independent, I argue that if strict determinism falls, it takes reductionism down with it. Turns out a little bit of randomness is the key to escaping Flatland.

Further reading:

Mitchell KJ. Does Neuroscience Leave Room for Free Will?. Trends Neurosci. 2018;41(9):573-576. doi:10.1016/j.tins.2018.05.008

Sara Imari Walker (2014) Top-down causation and the rise of information in the emergence of life. Information 2014, 5(3), 424-439; doi:10.3390/info5030424

Erik P. Hoel, Larissa Albantakis, Giulio Tononi. Quantifying causal emergence.

PNAS Dec 2013, 110 (49) 19790-19795; DOI: 10.1073/pnas.1314922110

Krakauer D, Bertschinger N, Olbrich E, Flack JC, Ay N. The information theory of individuality. Theory Biosci. 2020;139(2):209-223. doi:10.1007/s12064-020-00313-7

Noble R, Noble D. Harnessing stochasticity: How do organisms make choices?. Chaos. 2018;28(10):106309. doi:10.1063/1.5039668

How much innate knowledge can the genome encode?

noreply@blogger.com (Kevin Mitchell) — Sat, 04 Jan 2020 14:47:00 +0000

In a recent debate between Gary Marcus and Yoshua Bengio about the future of Artificial Intelligence, the question came up of how much information the genome can encode. This relates to the idea of how much innate or prior “knowledge” human beings are really born with, versus what we learn through experience. This is a hot topic in AI these days as people debate how much prior knowledge needs to be pre-wired into AI systems, in order to get them to achieve something more akin to natural intelligence.

Bengio (like Yann leCun) argues for putting as little prior knowledge into the system as we can get away with – mainly in the form of meta-learning rules, rather than specific details about specific things in the environment – such that the system that emerges through deep learning from the data supplied to it will be maximally capable of generalisation. (In his view, more detailed priors give a more specialised, but a more limited and possibly more biased machine). Marcus argues for more prior knowledge as a scaffold on which to efficiently build new learning. He points out that this is exactly how evolution has worked – packing that kind of knowledge into the genomes of different species.

Of course, the question is how that works. How does information in the genome lead to pre-wiring of the nervous system in a way that automatically connects external stimuli to internal values or associations?

In the course of a brief discussion on this point, Bengio argued that there is not, in fact, enough information in the genome to encode much prior knowledge: “…there’s lots of room in the genome but clearly not enough to encode the details of what your brain is doing. So, it has to be that learning is explaining the vast majority of the actual computation done in the brain, just by counting arguments: twenty thousand-odd genes, with a hundred billion neurons and a thousand times more connections”.

Marcus responded with a reference to his excellent book “The Birth of the Mind”, in which he refutes this “genome shortage argument”, which he suggests people take as implying that most of our knowledge is learned and that there simply isn’t enough information in the genome to specify a lot of priors.

Unfortunately, their discussion veered off this topic pretty quickly, but it’s a very interesting question and, in my view, the way that Bengio phrased it – as a simple numerical argument – completely misconceives the way in which the kind of information we are after is encoded in the genome.

He is not alone. This discussion harks back to an issue that exercised many people in the biological community when the Human Genome Project completed the first draft of the human genome and it turned out that we had far fewer genes than had previously been thought. Previous estimates had put humans at about 100,000 genes (each one coding for a different protein), based on extrapolation from a variety of kinds of data, the details of which are not really important. When it turned out we have only about 20,000 genes, there were gasps of horror, as this number was not even twice as many as lowly fruitflies or even vastly simpler nematodes, which each have ~15,000 genes.

Nematodes only have ~1,000 cells, ffs! If we are so complex (and we are, if nothing else, stubbornly impressed with ourselves), how can we have only on the same order of genes as a crappy little worm that doesn’t even have a brain? And, as Bengio and others argue, if we have only this limited number of genes, how can we encode any kind of sophisticated priors in the connectivity of our brains, which is mathematically-speaking vastly more complicated?

Counting the wrong things

To me, this kind of thinking misconceives entirely the way that developmental information is encoded in the genome. It’s not the number of genes that matters, it’s the way that you use them. From a developmental perspective, the number of proteins encoded is not the measure of complexity. The human genome encodes almost exactly the same set of proteins as every other mammal, with few meaningful differences in the actual amino acid sequences. In fact, the biochemical sequence and functions of most proteins are in many cases conserved over much larger evolutionary distances – from insects to mammals, for example.

As one famous experiment illustrates (and many others have found) it is often possible to substitute a fly version of a protein with the mouse version and have it function perfectly normally. The experiment I’m thinking of was looking at a gene called eyeless in flies (because mutant flies lacking the gene have no eyes) and Pax6 in mice. Eyeless/Pax6 is known as a “master regulator” gene – it encodes a transcription factor that regulates a cascade of other genes necessary for eye development, in both flies and mice.

If you drive expression of the Eyeless protein in parts of the developing fly that give rise to other appendages like antennae, legs, or wings, you can convert those appendages into eyes as well. Amazingly, if you express the mouse version, Pax6, in these tissues, the same thing happens – you form extra ectopic eyes (fly eyes, not mouse eyes). So, the differences in amino acid sequence between the proteins encoded by eyeless and Pax6 don’t make much of a difference to its function. What does make a difference is the context in which it is expressed – the network of genes that the protein regulates in each species and the subsequent cascading developmental trajectory.

Expression of eyeless in other tissues drives formation of ectopic eyes

This relies not on the protein-coding sequences, but on the so-called regulatory sequences of the target genes. These are short stretches of DNA, often adjacent to the protein-coding bits, which act as binding sites for other proteins (transcription factors or chromatin regulators), which control when and where a protein is expressed – in which cell types and at what levels.

From Innate

These regulatory elements typically work in a modular fashion and are much more evolvable than the sequences of the regulatory proteins. The latter are highly functionally constrained because each such protein regulates many targets and any change in its sequence will have many, diverse effects. (This is probably why they work the same even across distantly related species). By contrast, changing the binding site for one protein that regulates gene A will change only that aspect of the expression pattern of gene A which that protein regulates, leaving all other aspects and all other genes unchanged.

The complexity of those regulatory sequences is not at all captured by the number of genes. Indeed, it is not even captured by the number of such elements themselves, as they interact in combinatorial and highly context-dependent ways, in cooperation or competition with different sets of factors in different cell types and at different stages of development. (For example, Pax6 in mammals is also involved in patterning the developing neocortex and in specifying a number of different cell types in the developing spinal cord, in combination with other transcription factors). The complexity of the output thus should not be expected to scale linearly with the number of genetic elements.

Being amazed that you can make a human with only 20,000 genes is thus like being amazed that Shakespeare could write all those plays with only 26 letters. It’s totally missing where the actual, meaningful information is and how it is decoded.

How do we measure information?

Part of the problem is in thinking of Shannon information as the only measure of information content that is really scientific (nicely described in James Gleick’s “The Information”). Shannon information (named after Claude Shannon) simply relates to the efficiency of encoding information for signal transmission. It is measured in bits, which correspond to how many “yes or no” questions you would have to ask to derive the original message that you want to transmit.

Importantly, this does not take into account at all what the message means. Indeed, a purely random sequence of letters has greater Shannon information than a string of words that make up a sentence, because the words and the sentence have some higher-order patterns in them (like statistics of letters that typically follow each other, such as a “u” following a “q”), which can be used to compress the message. A random sequence has no such patterns and thus cannot be compressed.

Thinking in those terms naturally leads to the kind of “counting arguments” that Bengio makes. These seem to take each gene as a bit of information, and ask whether there are enough such bits to specify all the bits in the brain, usually taken as the number of connections. Obviously the answer is there are not enough such bits. (There aren’t even enough bits if you take individual bases of the genome as your units of information).

But the genome is not a blueprint. Bits of the genome do not correspond to bits of the body (or bits of the brain). The genome is much more like a program or an algorithm. And computer scientists have a much better measure to get at how complex that program is, which is known as algorithmic complexity (or Kolmogorov complexity). This is the length of the shortest computer program that can produce the output in question.

How complex is your code?

What we really would like to know is the Kolmogorov complexity of the human brain – how complex does the developmental program have to be to produce it? (More specifically, to get it to self-assemble, given the right starting conditions in a fertilised egg). And, most germane for this discussion, how complex would that program have to be to specify lots of innate priors?

Speaking as a developmental neurobiologist, if we knew the answers to those questions, we’d be done. We can’t figure out or quantify how complex the developmental program is until we know what the developmental program is (*but see note from Tony Zador, below). In fairness, we know a lot – the field is in some ways quite mature, at least in terms of major principles by which the brain self-assembles based on genomic instructions. But I think it’s also fair to say that we do not have a complete understanding of the molecular logic underlying the guidance of the projections and the formation of synaptic connections of even a single neuron (like this one, for example) in any species. What we can say is that that logic is highly combinatorial, meaning you can get a lot of complexity out of a limited set of molecules.

And we are only beginning to understand how instinct and innate preferences and behaviors are wired into the circuitry of the brain. There are growing numbers of examples of subtle genetic differences that lead to specific differences in neural circuitry that explain differences in behaviour. Some of these relate to differences in behaviour between closely related species (such as monogamy vs polygamy, or geometry of tunnel building). But probably the best-studied are sex differences within species, where a subtle genetic difference between males and females (the activity of the Sxl gene in flies or the presence of the SRY gene in mammals, for example) shifts the developmental trajectory of certain circuits, and pre-wires preferences and behaviours.

It’s an exciting time for this kind of research, with lots of amazing new tools being focused on fascinating biological questions in all kinds of species, but really this work is only just beginning. Moreover, it typically focuses on how a difference in the genome leads to a difference in circuitry and innate behaviour. It doesn’t explain the full program required to specify all of the circuitry on which any such behaviour is built.

So, we simply can’t currently answer the question of how complex the developmental program is and how much innate knowledge or behaviour it could really encode because we just don’t know enough about how the program works or how innate knowledge and behaviour is wired into the brain. But I think we know enough to know that the number of genes is not the number we’re after.

In fact, I’m not sure there is a number that could capture what we’re after. I understand the urge to quantify the information in the genome in some way, but it assumes there is some kind of linear scale. For signal compression (Shannon information) or algorithmic length (Kolmogorov complexity), you can generate such a number and use it to compare different messages or programs. I don’t know that that’s the case for the complexity of the brain or the complexity of the genome.

I’m willing to say that a human brain is more complex than that of a nematode and the human genome probably has correspondingly more information in it than the nematode genome. But does the human genome have more information than the mouse genome or the chimp genome? Or is it just different information – qualitatively, but not quantitatively distinct? Would the Kolmogorov complexity differ between the mouse genome and the human genome? Not much, I’d wager, and I’m not sure what you’d learn by quantifying it.

The real problem is that none of those measures capture what the information means, because the meaning does not inhere wholly in the message – it depends on who is reading it and what else they know. For example, here is a (very) short story written by Ernest Hemingway:

For sale: Baby shoes, never worn.

It doesn’t have much Shannon information and you could write a very brief program to reproduce it. But it’s freighted with meaning for most readers.

What does it all mean?

So maybe we should be asking: what is the meaning encoded in the genome and how is that meaning interpreted and decoded and ultimately realised? The impressive thing is that the interpretation is done by the products of the genome itself, starting with the proteins expressed in the egg, and then with the proteins that very quickly come to be produced from the zygotic genome.

The only thing physical property the genome has to work with to encode anything is stickiness. People often say that DNA is a chemically inert molecule that doesn’t do anything by itself and this is accurate in the sense that it does not tend to chemically react with or form covalent bonds with other molecules. But it is functional in a different way – it is an adsorption catalyst. It is a surface that promotes the interaction of other molecules by bringing them in to close proximity.

The specific sequence of any stretch of DNA determines which proteins will bind to it and with what affinity. If those proteins are transcription factors or chromatin proteins then their binding may regulate the expression of a nearby gene, as discussed above, by increasing or decreasing the likelihood that the enzyme RNA polymerase will bind to the gene and produce a messenger RNA molecule. And, in turn, the mRNA acts as an adsorption catalyst, bringing transfer RNAs carrying specific amino acids into adjacency so that a peptide bond can be formed between them, eventually forming a specific protein. It all starts with stickiness.

Of course, what unfolds after that – after the maternally deposited proteins bind to the genome in the fertilised egg and the new zygote starts making its own proteins – seems almost miraculous. Those patterns of biochemical affinity embody a complex and dynamic set of feedback and feedforward loops, coordinating profiles of gene expression, breaking symmetries, driving cellular differentiation and patterning of the embryo. And the biochemical activities and affinities of the proteins produced then control morphogenesis, cell migration, and, in the developing nervous system, the extension and guidance of projections, and tendencies to make synaptic connections with other cell types.

So, the developing organism interprets its own genome as it self-assembles. Pretty astounding, but that’s life (literally). However, the ultimate interpreter of the meaning in the genome is natural selection. The resultant organism has to be made with the right specifications, within the right operating range so as to survive and reproduce, given the right environmental conditions.

You don’t need to specify where every synapse is to get the machine to work right. You just need to specify roughly the numbers of different cell types, their relative positions, the other types of neurons they tend to connect to, etc. The job of building the brain is accomplished statistically, and, crucially, probabilistically. This is why there is lots of variation in brain structure and function even between monozygotic twins and why intrinsic developmental variation is such a crucial (and overlooked) source of differences in people’s psychology and behaviour (the subject of my book Innate).

Back to A.I.

Okay, so, where does this leave the question we started with? Is there enough information in the genome to specify lots of innate priors? Based on what I’ve said above, I don’t think the number of genes in the genome places a limit on this or is even an informative number. While I have some sympathy with efforts to define what Tony Zador refers to as a “genomic bottleneck”, I’m not convinced that quantifying it will be in any way straightforward or necessarily useful.

It’s certainly true that there isn’t enough information in the genome to specify the precise outcome of neural development, in terms of the number and position and connectivity of every neuron in the brain. The genome only encodes a set of mindless biochemical rules that, when played out across the dynamic self-organising system of the developing embryo, lead to an outcome that is within a range of operational parameters defined by natural selection. But there’s plenty of scope for those operational parameters to include all kinds of things we would recognise as innate priors. And there is plenty of evidence across many species that many different innate priors are indeed pre-wired into the nervous system based on instructions in the genome.

For A.I., it still may be best to try and keep such priors to a minimum, to make a general-purpose learning machine. On the other hand, making a true A.I. – something that qualifies as an agent – may require building something that has to do more than solve some specific computational problems in splendid isolation. If it has to be embodied and get around in the world, it may need as much help as it can get.

*[Note from Tony Zador: The length of the genome, which in humans is around 3 billion letters, represents an upper bound on the Kolmogorov complexity. Only a fraction of that carries functional information, however, so the upper KC value may be quite a bit lower than that].