Bayesian Just-So Stories in Psychology and Neuroscience (+ Commentary + Response below) | Bowers & Davis, 2012

Bayesian theories in psychology and neuroscience (according to the authors): minds and brains are (near) optimal in solving a wide range of tasks.

This article argues in favor of non-Bayesian accounts.

  1. Weak empirical evidence for Bayesian theories in Psychology
    • Manipulation of priors, likelihoods, utility functions in order to fit the data can be arbitrary.
    • So these theories are non-falsifiable
    • Bayesian theories are rarely better in predicting data that simpler non-Bayesian theories.
  2. Also weak empirical evidence in neuroscience
  3. Challenge to the theoretical framework of Bayesian theory in cognitive science
    • In specific against the idea that psychological theories should be constrained by what the mind ought to do
    • I.e. it is formulated on the computational level rather than the algorithmic one.
    • Instead of biological, evolutionary, processing constraints/considerations.
    • "Not only does this 'rational' approach to cognition lead to underconstrained theories, it dissociates theories of cognition from a wide range of empirical findings in psychology and neuroscience"

Overall claim: we can often speak of Bayesian "just so" stories because they offer "mathematical analyses of cognition that can be used to explain almost any behavior as optimal" (389).

In our view, the flexibility of Bayesian models, coupled with the common failure to contrast Bayesian and non-Bayesian accounts of performance, has led to a collection of Bayesian "just so" theories in psychology and neuroscience: sophisticated statistical analyses that can be used to explain almost any behavior as (near) optimal. If the data had turned out otherwise, a different Bayesian theory would have been carried out to justify the same conclusion, that is, that the mind and brain support near-optimal performance. (390)

Other accounts always suggested "good enough" solutions, with the mistakes still being quite salient.

  • (Adaptive) heuristics in high-level reasoning.
  • Perception as a "bag of tricks"
  • Evolution produces systems that satisfice (good enough) or meliorize (better than alternatives).

The Bayesian approach claims evolution gave us brains that are great at learning and exploiting the statistics of the environment in a way that is close to optimal.

N.B. article does not argue against Bayesian statistics itself, or for evaluation, but as a description of mind and brain.

Some questions

Is Bayesianism really pitted against heuristic approaches?

On page 389 Bowers et al. contrast the Bayesian approach to heuristic approaches.

The Bayesian approach, by contrast, appears to claim that evolution has endowed us with brains that are exquisitely good at learning and exploiting the statistics of the environment, such that performance is close to optimal. (389)

Is this however a fair opposition?

Is the Bayesian Brain committed to "storing" probabilities

Kwisthout in response to p. 393:

Bowers and Davis claim that a Bayesian algorithm must store priors in the form of probability distributions, compute estimates of likelihoods etc. Most neuroscientists that uphold the Bayesian Brain hypothesis will not agree with this claim. For example, a Bolzmann machine (or a spiking neural net, to make it more biologically plausible) inherently represents by its emergent firing behavior a probability distribution on which one can perform inference by clamping nodes to specific outputs. There is nowhere any explicit storage of priors, and the distribution can only be sampled from (every readout of the neurons is a sample from the distribution), but the system as a whole is clearly capable of approximate Bayesian inference.

Bayesian theory in psychology and neuroscience

  • Brain does Bayesian statistics (approximately)
  • because it is the optimal method for estimating probabilities of unique events
    • vs frequentist statistics.

Complications with Bayesian statistics w.r.t. cognition.

  1. People are quite bad at taking into account prior and estimating it and or likelihood
  2. Prior and likelihood are in practice distributions rather than point estimates, introducing uncertainty
  3. Posterior distribution does not yet indicate what the optimal decision is
    • So for optimality theory, often a secondary utility function is taken into account that integrates other contextual factors
    • "the estimation of utility functions in psychological models is not straightforward, as both the variables that affect utility and the shape of the function relating these variables to utility tend to be unknown." (393)

What does optimality mean here?

In the Bayesian context: optimal given noisy input data!

Noise results from a combination of uncertainty in the environment and suboptimal human "design".

In sum, Bayesian statistics provide a method of computing the posterior probability of a hypothesis, which provides the best way to update a prior belief given new evidence. Bayesian decision theory defines how our beliefs should be combined with our objectives to make optimal decisions given noisy data. (393)

Three types: How does the computational level relate to the algorithmic one?

  1. Extreme
  2. Methodological
  3. Theoretical

The authors argue there is a profound confusion between authors between the methodological and theoretical perspective. They also argue it is however very important to have this distinction clear, because it really changes how to test the merit of a Bayesian approach, over other alternative and non-Bayesian theories. (394)

Extreme: Psychology should concerns only with computational level theories. ("How" the mind computes, is an intractable question). This viewpoint dismisses what most cognitive scientists find the most interesting.

Methodological:

  • Bayesian model is simply a tool: "measure of optimal behavior that serves as a benchmark for actual performance" (393).
  • Not committed to any algorithmic-level theory or specific theory of mind. Irregardless of whether full-blown Bayesian implementation or heuristic is used on the algorithmic level.

From this perspective, the striking result is how often human performance is near optimal, and this is considered useful for constraining a theory (whatever algorithm the mind uses, it must support behavior that approximates optimal performance). (393)

Theoretical:

agree that Bayesian models developed at the computational level constitute a useful method for constraining theories at the algorithmic level (consistent with the methodological Bayesian approach), but claim that the mind carries out or approximates Bayesian computations at the algorithmic level, in some unspecified way. (393)

  • And also, potentially, at the implementational level.
  • But again, not committed to one specific algorithm/implementation, apart from that it must be Bayesian or approximately so.
  • Challenge is to approximate Bayesian priors and likelihood etc. in a psychologically plausible way (on the algorithmic level, since pure Bayesian is computationally intractable).
  • This may require heuristic approach, but there is a difference with heuristic models w.r.t. the claims they make:

the output of a good heuristic model might approximate the decisions of a Bayesian model in a given context, but it will not approximate the underlying processes. (394)

Bayesian theories are flexible enough to account for any pattern of results

Overall claim: " there are too many arbitrary ways that priors, likelihoods, utility functions, etc., can be altered in a Bayesian theory post hoc" (394).

See paper for examples:

  • Speed perception illusion -> outcome depends on arbitrary prior choice
  • Word identification
  • High-level cognition

High-level cognition is an interesting case, because the work of Kahneman etc. show exactly that humans are quite poor decision makers, susceptible to various irrational fallacies. There is a resurgence in Bayesian theories of optimal decision making, but the authors here state these rely on strong post hoc assumptions.

  • Explanation of behavior can be ad hoc explained by changing the utility function used
  • Purely Bayesian principles do often not sufficiently explain behavior, but require ancillary assumptions.
    • For example an algorithmic theory with processing assumptions.

Are Bayesian theories not falsifiable?

Main issue, non-falsifiability:

If explanatory failures in a Bayesian model can be minimized by attributing the performance in question to processes occurring at the algorithmic level and outside the scope of the computational theory, is it ever possible to falsify Bayesian models, situated at the computational level? And if the behavior of the model may differ when the specific processing assumptions of the algorithmic level are implemented, how much weight should be attached to the successes of the computational theory? And if the behavior of the model may differ when the specific processing assumptions of the algorithmic level are implemented, how much weight should be attached to the successes of the computational theory? Might not these "correct" predictions be modified when the algorithm is implemented? (397)

If you require ad-hoc introductions to explain the data, then your theory is descriptive, rather than normative / rational (397).

If Bayesian is to be normative, its priors and likelihood etc. need to be chosen independent of the observed data i.e. only with respect to the problem at hand.

Falsification

  1. Bayesian models are supposedly more constrained than their non-Bayesian counterparts because they have less free parameters
    • But this is not a fair comparison, the authors argue, because of the aforementioned freedom in choosing priors and likelihood and utility (which seems to be forgotten in this argument)
  2. Bayesian models less easily falsified because they make fewer predictions
    • "That is, models developed at a computational level do not make predictions about phenomena that can be directly attributed to the algorithmic or implementational level of analysis." (397and and )
    • E.g. how can you computationally and rationally explain the effects of brain lesions?
    • They don't, but as a consequence, Bayesian models do not make predictions here and accordingly cannot be falsified by these predictions.

Insight in the "Why" question?

Flexibility is considered as a selling point from another perspective: i.e. Bayesian theories can provide insight in the why question.

Type 1 answer:

If a Bayesian theory makes predictions based on rational analysis, and behavior corresponds to it, then the first answer to the why question is: behavior is what it is because it's close to optimal. I.e. apparently the rationality assumption of the Bayesian model holds if it indeed predicts the behavior.

Type 2 answer:

"People act in a certain way because they are rational and have specific assumptions about the world." (398)

Quote from Griffiths and Tenenbaum, 2006, p. 772:

Bayesian models require making the assumptions of a learner explicit. By exploring the implications of different assumptions, it becomes possible to explain many of the interesting and apparently inexplicable aspects of human reasoning"

Claim of authors: neither of these warrant a strong preference for Bayesian or non-Bayesian theories.

Problem with Type 1:

  • Analysis rarely strictly rational
  • Mainly because of all the required choices for prior, likelihood etc. in order to capture the data
  • Predictions are thus descriptive rather than normative.

Main point w.r.t. Type 2:

There is no real difference between making additional assumptions (and testing those) and for example an evolutionary perspective, that explains irrational behavior "in an adaptive manner by looking for assumptions that the brain carries from our evolutionary past" (399). So there is nothing particularly Bayesian about this type of answer.

Bayesian theories are rarely compared with alternative (Non-Bayesian) hypotheses

in order to provide some evidence in support of the claim that the mind relies on Bayesian-like algorithms, it is necessary to show that these algorithms do a better job than non-Bayesian models in accounting for human performance. However, this is rarely done. Instead, theoretical Bayesians generally take the successful predictions of a Bayesian model as support for their approach, ignoring the fact that alternative theories might provide an equally good fit to the data. (399)

From 399 to 403, many examples.

Confirmation bias

Key point: proponents of Bayesian models show confirmation bias: successful predictions are taken to be support for the model, without considering that other non-Bayesian accounts might account for the data equally well or even better.

How would we distinguish Bayesian and non-Bayesian models?

Problem: distinguishing factors (nonobvious or counterintuitive predictions, better than heuristic models) are likely to be on the algorithmic level.

The nonobvious and counterintuitive predictions of a psychological model are typically derived from nonfunctional constraints found at the algorithmic level, a level of descriptions that most Bayesian models avoid. (403)

Bayesian models tend to receive credit for relatively trivial predictions, which could be derived from any theory that assumes performance is adaptive, or else are modified with post hoc assumptions when the data do not follow what would be predicted on the basis of the rational analysis alone. (403)

Neuro evidence for Bayesian theories is weak

In other words, we've discussed the computational and algorithmic level so far, and now turn to the implementational level.

Bayesian coding hypothesis: "populations of neurons represent uncertainty in the form of probability distributions and perform (or approximate) Bayesian computations in order to make optimal decisions." (403)

The authors say: indeed, neuroscientific evidence for Bayesian coding hypothesis would support the theoretical Bayesian perspective. They point out however, that the psychological literature (so behavioral evidence) is often cited as support for the neuro evidence.

For a strong case, the question is thus: what nonbehavioral evidence is there that neurons compute using probability distributions?

Kwisthout: Nowadays there is much more neuroscientific evidence e.g. in the predictive processing literature.

Does neural noise necessitate Bayesian inference?

Neural firing is noisy, i.e. for the same inputs you usually have very different neural responses.

A key insight of Ma, Beck, Lathum, & Pouget (2006) is that if neurons show near-Poisson variability, then Bayesian inference can be implemented through a simple linear combination of populations of neural activity. A fundamental property of a Poisson response is that the variance and the mean are the same, and a number of experimenters have found that the ratio of the variance to the mean spike count to range from about 1 to 1.5 (e.g., Shadlen & Newsome, 1998) the so-called Fano factor.

  • Sometimes the Fano factor approaches 0, suggesting the absence of noise. Meaning: this discredits a model (e.g. Bayesian) that assumes noise as an inevibility and also takes its main strength from it.
    • Of course, in any case populations of neurons may compute Bayesian-like.
    • The point remains the same: the selling point of the Bayesian approach is dealing with noise plus the assumption that noise is an issue on all levels of processing.

Evidence for neural coding hypothesis

The main claim here is that although there are mathematical models that describe how Bayesian models can be implemented in populations of neurons (given a certain noise level), this does not yet mean that neurons in fact do Bayesian computations.

In this section they discuss several examples, concluding that "it is difficult to relate neural firing to the Bayesian coding hypothesis." (405)

N.B. we are talking of the theoretical strand of Bayesianism:

the Bayesian coding hypothesis is relevant only to the theoretical Bayesian perspective. That is, it makes sense to consider the neuroscientific evidence if the claim is that various cognitive, perceptual, and motor systems rely on Bayesian-like algorithms, whereas neuroscience is irrelevant if Bayesian models are merely a yardstick to measure behavior (with no claim regarding the processes, let alone the neural mechanisms, that underpin performance).

Conceptual problems

Bayesian theories in psychology tend to adopt the rational analysis methodology of Anderson (1991), according to which the important constraints on cognition come from the nature of the problem and the information available from the organism (the environment). (405)

This assumes what Dennett calls a "triumphant cascade" through the levels of Marr! That is, this rational approach departs from the computational constraints imposed by the problem, which in turn trickles down as the main constraints on the algorithmic level and so on.

This cascading approach then indeed downplays the effect of biological and evolutionary insights (algorithmic?) on constraining psychology (high-level, computational).

The authors have pointed out that Bayesian models do need to make (arbitrary) assumptions in order to fit behavioral data/performance and add that these additional constraints and assumptions should instead come from biological considerations, both on the algorithmic and implementational level.

Examples from biology and/or evolution. I think the point w.r.t. evolution is interesting.

Evolution is geared towards good enough (satisficing) or better than other (meliorizing) solutions, rather than optimal ones. This process results in solutions that no engineer would come up based only on a given problem and rational analysis.

Not only would no engineer come up with this solution, but neither would a rational analysis constrained only by the environment and the task at hand (...). More generally, we would question the commitment to optimal decision making given that the human phenotype is nonoptimal in so many respects. And even if we were optimal, given a poor design, the standard Bayesian approach is incomplete, as it does not provide an explanation for why the design of many systems is so poor from an engineering point of view. Presumably, evolutionary inertia and physiology will provide the key insights here. (407)

When Bayesian models do consider biological and evolutionary constraints

  • Usually applies to low-level vision (because here there's close to ideal performance?)
  • these Bayesians are usually methodological rather than theoretical
  • I.e. they usually support a Bayesian model as a benchmark for evaluating theories

Bayesian as benchmarking:

Under these conditions, a specific algorithmic theory can be rejected because it does not capture human performance. The mismatch in accounting for data is sufficient to rule out the model [... there's a tab here that seems faulty] the extra step of showing that it does not support optimal performance is superfluous. (407)

Other point of criticism:

That is, the findings are consistent with most types of algorithmic theories that have been advanced independently of any rational analysis. (407)

Meaning that...

In this case, the rational analysis has not strongly constrained the algorithmic level, and we are not aware of any examples in which a Bayesian theory achieved the goal of constraining algorithmic (mechanistic) theories above and beyond the constraints imposed by the data themselves. (407)

Positive features of Bayesian models?

What are key contributions of the Bayesian approach?

  • Emphasized importance of environment
  • and importance of "why" questions
    • the criticism on other approaches is that often they "can be characterized as an exercise in curve fitting" (408)
  • They have attention for top-down constraints

But the authors again argue that nothing about these points is inherently Bayesian; they should inform good algorithmic level theories too. The merit of the Bayesian approach is merely that they have renewed attention for these issues.

Summary

In our view, the methodological Bayesian approach does not provide a useful tool for constraining theories, and the theoretical Bayesian perspective is unsupported. More generally, we claim that it is a mistake to develop models that optimize at a computational level given that the mind and brain satisfice, and given that many of the key constraints for models will be found at the algorithmic and implementational levels. Our main thesis is that Bayesian modeling, both in practice and principle, is a misguided approach to studying the mind and brain. (408)

Conclusion

Funny approach. The authors are extremely critical of Bayesian theories based on evidence that others take as arguments for Bayesianism. Interesting.

So what the authors do is frame this problem in terms of Bayes' rule, in order "to evaluate the probability of the hypothesis that humans are optimal Bayesian estimators, given the available evidence." (409)

So we have H_optimal.

Claims about why the posterior of Bayesian proponants is too high:

  1. prior is overestimated
    • should be lower based on much evidence of flaws in human reasoning and perception, and biological, evolutionary and behavioral constraints.
  2. Likelihood is overestimated
    • due to lack of constraint in Bayesian theories (purely rational considerations)
    • researchers have too much freedom in ad hoc choosing hypothesis that optimize the likelihood (form of overfitting, basically)
  3. Underestimation of the marginal probability of the evidence
    • Write out the expanded formula; you'll see that it includes probabilities about when the hypothesis is not true, e.g. of non-Bayesian theories
    • As the authors kept pointing out: often other non-Bayesian theories can also explain the evidence

As a result, Bayesian theories overestimate the probability that humans are near-optimal estimators.

Result: just-so stories "that can be used to explain almost any behavior as optimal" (410)

This in turn also invalidates the use of Bayesian theories as "benchmarking":

Both Bayesian and non-Bayesian theories can offer answers to why questions, and the relevant metric to evaluate the success of a theory is actual behavior, not optimal performance by some criterion. (410)

Comment from Griffiths, Chater, Norris, Pouget

Bowers and Davis show an overly pessimistic account based on handpicked examples, and under-represent the merits of the Bayesian approach. There are specifically six misconceptions in their pessimistic representation.

Misconception 1: Goal of Bayesian modeling is to show that people are optimal

"Rather, Bayesian models typically aim to provide explanations of human behavior." (415)

"Comparing human behavior to optimal solutions does not imply a belief that pople are actually computing those optimal solutions." (415)

Bayesian models are not just empirical hypotheses.

The main difference from other kinds of models are that (a) the content of the Bayesian model is a proposal about the problem that people are solving rather than a characterization of the mechanisms by which they might be solving it, and (b) the style of explanation is teleological rather than mechanistic. (416)

The latter point explains where the confusion may come from. Comparison with 'optimal models' is teleological, because it tries to elucidate "why" certain behavior is the way it is rather than describing it on the "what" level. (The "why" level is sometimes called the 0-th level of Marr).

Bayesian models are thus means of generating empirical hypotheses.

And not 1) claim about optimality of human behavior or 2) claims about what goes on "inside people's heads"

The authors of the comment would understand themselves as "methodological Bayesians" but also claim that they do not believe the theoretical position actually exists.

2 Bayesian models are unfalsifiable and overly flexible

We should distinguish

  1. Model: proposes account for a phenomenon and makes assumptions to this end
  2. Framework: provides "general perspective and a set of tools for making models." (416)

Models should be falsifiable but frameworks are usually not. They should instead be measured by their ability to generate useful models.

Specific Bayesian models are falsifiable. The point about ad-hoc assumptions also holds for non-Bayesian models; you can often save a hypothesis with ad-hoc patches.

The claim about that the Bayesian framework is not falsifiable thus makes a conceptual mistake.

Note that this argument does rely on the previous point. Falsifying the framework would be an issue if it would include a claim about the optimality of human behavior.

Another point was that there was too much flexibility in choosing priors etc. But each computational framework needs to deal with this issue and justify the choice, e.g. in neural network architecture. The quality of these choices is in turn tested by experiments to test made predictions.

Note that if you have a model with too many degrees of freedom, it has very little explanatory value. So this is inherently punished.

The big counter-argument is: but Bayesianism makes those degrees of freedom salient (making it an easy target). Other models have them as well, just implicitly, is the claim.

In other words, in Bayesianism these assumptions are principled.

3. Bayesian Modelers believe that Bayesian Processes underlie human cognition

The theoretical Bayesian also believes that the mind does approximate Bayesian computations at the algorithmic level.

The response is that these type of Bayesians may not exist. Bayesian models of cognition work on Marr's computational level without making direct claims about cognitive processes (the algorithmic level).

When explaining cognition, it is easier to confuse levels of analysis, given that Bayes's rule can be viewed as a procedure for updating beliefs as well as a mathematical solution for what those beliefs should be. Nonetheless, most Bayesian models appeal to the mathematical solution rather than the algorithmic procedure. (417)

The main point is that on the computational level the constraint on the algorithmic level is, that it must approximate Bayesian inference etc. But this does not imply in any way that this algorithm would be Bayesian.

Example of how the approximation itself does not inherently need to be Bayesian (cf. the casual psychological notion of "memory"):

The idea behind this approximation is that people store examples of past events in memory, which act as samples from the prior, and then activate these stored exemplars based on their similarity to observed data, which acts like the likelihood function. (417)

An interesting remark from the TFCA course: if you look close enough on what happens when we use a computer to multiply two numbers, at some point we'll stop seeing "multiplication". And yet, we can safely agree that what the computer is doing is in fact multiplication.

Note that the Bowers paper seems to conflate approximation with heuristic, but there is a clear difference: approximations have to guarantee some lower error bound on the output.

4. Bayesian models are never compared to other models

Main argument: it is not clear how to compare models on different Marr's levels. They may simply not be competing, requiring one is accepted over the other. They can both be valid.

5. Probabilistic population codes require noisy neurons

We should get clear what would be the question of Bayes on the neural level. It is claimed in the response that the central questions are:

  • "Do neurons encode single values, or do they also encode information about the certainty of the stimulus?" (418)
  • "Can they represent probability distributions (or likelihood functions) even when they have multiple peaks?" (418))
  • "do neural computatoins take into account these probability distributions?"

Is this theory proven wrong if neurons turn out to be nearly deterministic devices? The answer is clearly no. It is critical not to confuse variability due to the external world, which is the focus of Ma et al. (2006), with variability due to internal neuronal noise. Whether neurons are stochastic is irrelevant to the probabilistic population code theory. (418)

6. Bayesian Modelers do not believe constraints from biology and evolution are important

Basically, the argument here is that Bowers and David confuse a function-first approach with a function-only approach. Marr himself also said starting at the computational level is appropriate and effective. But ultimately the goal is an integrated theory that integrates all three levels. For this, biology and evolution are highly relevant.

As counterweight: merits of the Bayesian approach

Universal laws

Teleological explanation addresses the "why" question, but are especially interesting because they concern the foundations for "universal laws of cognition", for any type of intelligent organism.

This seems to be a very big claim to me. Example: Shepard's universal law of generalization.

From wiki at https://psychology.wikia.org/wiki/Universal_law_of_generalization:

"The universal law of generalization is a theory of cognition originally posited by Roger Shepard. According to it, the probability that a response to one stimulus will be generalized to another will be a function of the distance between the two stimuli. "Generalization" in this case is measured by means of confusion error, while the use of "distance" depends on the assumption that stimuli will be compared in some kind of psychological space (the latter being typical of Shepard's work)."

Deriving general predictions

The Bayesian framework of inference provides a general framework for learning and memory. Discoveries on this general level may lead to predictions in a vast array of different fields related to learning and memory.

Understanding why particular mechanisms work