Lecture 5: Bayesianism in Cognitive Science and Neuroscience

This lecture transitions from "Bayesianism" as a more philosophical framework towards more concrete implementations inspired by this framework.

No exam questions on that chapter in lecture notes on stochastic computations. We have to skip that due to Corona etc.

Recap

We have not seen each other for 5 or 6 weeks, so we recap is in order.

We have discussed:

  • Mars Rover case study and explainability
  • Philosophical issues: makes sense of sensory information, inference to the best explanation, abduction.
  • Frame problem
  • These issues with symbolic AI do not seem to be applicable to sub-symbolic AI, but is reintroduced once explanation to the user becomes important
  • Problems with the Bayesian approach
  • What is a probability?
  • Objective vs subjective Bayesianism?
  • etc.

The Bayesian turn in cognitive science

Johan Kwisthout, Todd Wareham, Iris van Rooij (2011). Bayesian Intractability is not an Ailment that Approximation can Cure. Cognitive Science, 35(5), 779-784.

Describes a huge increase in papers about Bayesian methods in proceedings of the Cognitive Science Society.

Authors such as Josh Tennenbaum, Griffits really boosted the field, not only as a methodological tool, but also as relating to a specific framework describing cognition.

Two trends/discussions:

  1. The Bowers and Davis paper we read concerns a debate about the specific role of Bayesianism for cognitive science. General theory.

  2. A similar debate is going on in relation to cognitive neuroscience under the term "The Bayesian Brain": assumption that the brain somehow integrates and unifies new sensory information in a (more or less) Bayesian way. A related hypothesis is the predictive processing hypothesis, stating that the brain tries to minimize prediction error in perception.

To Bayes or not to Bayes

It is really difficult to falsify the Bayesian cognitive "theory", but that's really because its not a theory but a framework. It should instead be tested by the models it allows and generates, that should be falsifiable and should make specific predictions.

As a framework, Bayesianism provides axioms of cognition.

Debate in the Bowers and Davis paper: "Are we as humans optimal? And does Bayesianism commit to this 'optimality' claim?"

The "teleological" approach: faces with a problem, the optimal way of solving it is with Bayesian reasoning to reach a posterior distribution; optimal combination of uncertainty in inputs with our prior knowledge. Bayes can for example also express that visual input has a more narrow distribution than auditory input. Both are more effective when combined with prior distribution about possible locations of some sensory input (e.g. a bird does not live underwater).

This is what brains ought to do, if they want to solve this problem optimally.

Optimality claim w.r.t. brains: brains are trying to approximate this Bayesian optimal approach.

Paradoxically, we are very bad at estimating probabilities and combining priors and posteriors (point of critique, attacking some version of the optimality claim). So Bayesianism is not descriptive for actual human behavior. (Potential answer: it doesn't try to be)

Critique: Bayesian framework has too many degrees of freedom. For any type of dataset, there is some Bayesian model we can construct to fit the data.

This limits the usefulness of Bayesian framework as explanation of behavior.

Underlying these remarks: Do Bayesian computations require Bayesian implementations?

Potential counter: will you try to find evidence of "multiplication" in a computer when you measure machine code at the CPU level, or the logic-gates, or differential power changes? Where does "multiplication" happen?

We nevertheless know (we wrote e.g. Python code) that application is being performed.

Analogically, brain might do Bayesian computations with non-Bayesian mechanisms.

Then the question is, how are Bayesian computations realized? Approximation or with heuristics?

When does something approximate? You need some guaranteed lower-bound with respect to the true value you approximate.

Terminology

Bayesian cognitive science:

  • Cognitive systems behave as rational Bayesian agents, i.e. what they ought to maximize in order to optimally give a best possible prediction about some perception (in order to maximize behavioral success).
  • Provide a "golden standard" for what brains are trying to approximate.

Bayesian coding hypothesis:

  • Neuroscientific assumption that the brain encodes and computes with probability distributions (leaves open various implementations for encoding probabilities in the brain; e.g. whether it is encoded temporally).

Bayesian brain hypothesis:

  • In addition, the brain combines prior knowledge and sensory inputs in an approximately Bayesian manner.

Free energy principle:

  • axiom that biological systems resist disorder and thereby minimize prediction errors; where "prediction error" is a bit unhandy, its more the difference with the environment expected for homeostasis.

Predictive processing principle:

  • Combination of previously three

Q: what's the difference between Bayesian cognitive science and Bayesian brain?

A: Bayesian cognitive science does not make assumptions about the algorithmic and implementational level, only on the computational level. "Whatever" the brain does, it would be some approximation of Bayesian inference.

The Bayesian Brain takes this a step further and thinks that the brain operates in a roughly Bayesian way.

But they do not assume each other. The Bayesian brain does also not necessarily assume the cognitive approach that what is being approximates at the computational level is Bayesian. (Although this seems a bit unlikely).

Are humans optimal?

What is optimality in this context?

Are we searching for a unifying account (Bayesian model) or is optimality the result of an adaptive toolbox of heuristics? (Like the "neats" vs "scruffies" debate in AI)

Revisit the aforementioned paradox, often used as an argument against the Bayesian framework in cognition: Tversky & Kahnemans' work on estimating probabilities.

  • Wason selection task
  • Linda the feminist bank teller

Common point: how can it be that our brains approximate Bayesian inference while our conscious estimations are so biased?

Are our Bayesian inferences "cognitively penetrable"?

The framing of the rules matters a lot.

Wason selection example:

A, D, 3, 7. Rule: If there is an A on one side, there is a 3 on the other side.

If A, then 3. If not 3, then not A.

You need to check A and 7. (N.B. you do not choose D because one side always has digits; the other a letter).

Linda the bank teller

Which is more probable?

  1. Linda is a bank teller?
  2. Linda is a bank teller and in the feminist movement?

The second fits the pretext more, but can in no way be less probable than 1 because it is a conjuntion of 1. So if any statement is more probable, it has to be 1.

Do Bayesian models allow for too much freedom?

In theory: Jaynes' Maximum Entropy Principle states you should limit your distribution to only the available information, and maximize uncertainty otherwise.

In practice: priors are being fitted to the data for post hoc explanations.

What we measure in Bayesian experiments does not necessarily represent probabilities.

So in predictive processing we want to measure prediction error in the brain. But we do not measure this! In fact, we measure some proxy.

E.g. pupil dilation or increase in reaction time as a proxy for prediction error, but you are not directly measuring "in the brain" what you are interested in.

Approximation and heuristics

Intractable functions do not have tractable algorithms.

Heuristic:

You could then have a tractable heuristic. But the heuristic is not guaranteed to approximate within a lower bound, how do you show it is consistent with the computational level?

Main issue here is that because there is an inconsistency between the optimal output and the heuristic output, you have to make a choice: Would you then still maintain the computational level explanation? Or would you move the heuristic itself to the computational level?

Translated to the Bayesian Brain: if you consistently resort to using heuristic, maybe your computational level theory does not hold up too well.

Approximation:

Difference with heuristic is that it guarantees some bounds to the approximation of the optimal output. Already discussed.

This also means that the above issue does not arise.

Case study: analogy between multiplication and non-Bayesian implementation

  • Assume multiplication in high level programming language. At what point do you still refer to it as multiplication?
    • Compiled into assembly including mul operator? Yeah probably
    • Machine language?
    • Voltage pattern: notion of multiplication is gone.

Now make a parallel: what is needed to call a computation Bayesian?

Johan's position:

  • There need not be a any explicit storage of priors, likelihoods, or computational of posteriors in the brain to make the information processing "Bayesian"
  • We could encode hyperpriors as natural numbers and compute with them as if they are probabilities
  • Bolzmann machines or networks of spiking neurons represent a probablity distribution with firing patterns; distribution is only retrieved by sampling.
  • Just like there is no "multiplication" in the above case study, just so there does not have to be something specifically "Bayesian" in the implementation.

Free energy principle

Specific unifying principle in neuroscience, about what life itself and cortical activity as a subset, tries to accomplish: minimize entropy.

Minimizing "prediction error" between assumptions about the world and feedback from the world, where these assumptions allow the organism to live and prosper.

Prediction error is here thus (perceptual) feedback.

Prediction is just not something I "believe" to be the case, but something I "desire" to be the case.

You can't really prove or disprove it, but it has interesting relevance for justifying frameworks/models.

Predictive processing

In continuation of the free energy principle.

A particular implementation of the Bayesian ideas. More of a philosophical framework.

Three main components:

  1. Brain is a prediction machine
    • Brains continuously makes predictions about future sensory evidence based on its current best model of the causes of such evidence. And prior model.
  2. Bayesian brain combines prior knowledge with sensory evidence in a Bayesian way
  3. Hierarchical brain: "high level" information influences "low level" information and vice versa.

Brains are equipped with generative models that link perceptions with probable causes. Tries the minimize the prediction error between the expectation and the actual perception. (Change perception to better fit input?)

There are two main ways to fix "prediction errors":

  • Perceptual inference
  • Action inference

Whether the perception or action should be updated also depends on homeostasis/survival reasons. More of a philosophical debate perhaps.

Open questions:

  • How do these generative models updated or replaced?
  • And can brains actually integrate novel, "created" information?
  • How is this operationalized in the brain?
  • How does a generative model look like?
    • If you expect a hand movement, this concerns some expected trajectory with some dynamics, and noise describable with a Gaussian. But if it concerns an expectation that someone goes shopping, we cannot express this with a distribution. The level of detail here is different. What is the state space of these predictions? What is relevant for any given prediction? When crossing the street, I don't care whether a coming car is a Suzuki or Mercedes. The point is: we make predictions at whatever granularity/level of detail is relevant. We are surprised to find a Suzuki in a Mercedes dealership.