Inspired by this question: What are some of the drawbacks to probabilistic models of cognition?

I would like to know more about the biological plausibility of Bayesian models of cognition. Is there any neural evidence that rejects Bayesian models of cognition?

  • 2
    $\begingroup$ what are some papers you looked at already while thinking about this? $\endgroup$ Mar 27, 2012 at 17:50
  • 1
    $\begingroup$ I think one problem with this question is that bayesian models work on a completely different level of abstraction than neuroscience data. Any data can be interpreted as being optimal for something $\endgroup$
    – zergylord
    Mar 27, 2012 at 19:16
  • $\begingroup$ IMO this question needs a lot more fleshing out before it can be helpfully answered. $\endgroup$
    – Ben Brocka
    Mar 30, 2012 at 0:43
  • $\begingroup$ I edited your question title (and moved the original title to the body), hopefully this captures the spirit of what you were asking, if not feel free to rollback my edit. Also, you might be interested in this comment/answer $\endgroup$ Apr 3, 2012 at 19:22
  • $\begingroup$ At the Pillow Lab Blog you can find a brief discussion & summary of two different threads in the conversation about "Probabilistic Representations in the Brain" (with references). $\endgroup$
    – yep
    Apr 4, 2012 at 1:15

2 Answers 2


When performing certain tasks, people’s inferences approximate Bayesian inference to a remarkable degree. For example, when people receive both haptic and visual information about the size of an object, they combine this information in a manner that very closely resembles Bayesian inference, taking account of the uncertainties associated with the visual and haptic information (e.g., Ernst & Banks, 2002). This optimality can be observed in many perceptual (Knill & Pouget, 2004) and sensorimotor (Kording & Wolpert, 2004, 2006) tasks and across a range of information sources (e.g., including prior beliefs and multiple sensory inputs). These findings suggest that there must be biological mechanisms that either implement Bayesian inference or implement something that very closely resembles it.

At the same time, there is no consensus regarding how this is done. While there are proposals about how neural populations might perform Bayesian inference (e.g., Ma, Beck, Latham, & Pouget, 2006; Knill & Pouget, 2004; Kover & Bao, 2010), it is difficult to evaluate these proposals at present: the available neuroscientific evidence is quite limited. Moreover, because the most compelling evidence for Bayesian inference is limited to low-level perceptual processes, it is possible that higher-level inferences are implemented by biological mechanisms that do not perform Bayesian inference. Indeed, given the computational difficulty of Bayesian inference in general, it seems all but inevitable that many biological mechanisms will not implement Bayesian inference exactly.

In summary, some biological mechanisms must perform something like Bayesian inference, but researchers have only begun to explore how this happens, and the extent to which high-level perception and cognition rely on Bayesian computations remains unclear.


Knill, D. C., & Pouget, A. (2004). The Bayesian brain: the role of uncertainty in neural coding and computation. Trends in Neurosciences, 27 (12), 712-719. [pdf]

Kording, K. P., & Wolpert, D. M. (2004). Bayesian integration in sensorimotor learning. Nautre, 427, 244-247. [pdf]

Kording, K. P, & Wolpert, D. M. (2006). Bayesian decision theory in sensorimotor control. Trends in Cognitive Sciences, 10 (7), 319-326. [pdf]

Ma, W. J., Beck, J. M., Latham, P. E., & Pouget, A. (2006). Bayesian inference with probabilistic population codes. Nature Neuroscience, 9 (11), 1432-1438. [pdf]

Ernst, M. O., & Banks, M. S. (2002). Humans integrate visual and haptic information a statistically optimal fashion. Nature, 415, 429-433. [link]

  • $\begingroup$ Welcome to cogsci.SE! This is a great answer! Minor comment: could you edit in a reference section at the end to make finding the articles you cite a little bit easier? Thank you! $\endgroup$ Apr 4, 2012 at 0:16
  • $\begingroup$ Here is a relevant discussion on a suggested way of referencing papers. $\endgroup$
    – Steven Jeuris
    Apr 4, 2012 at 0:18
  • $\begingroup$ Further, I would love to see your comment about the computational difficulty of bayesian inference expanded into an answer for this question if you have the time and energy. $\endgroup$ Apr 4, 2012 at 0:21

For a fruitful discussion of this, which includes some useful links in the comments section, you might want to read this LessWrong post.

Whatever human decision making is at this moment in time, it is something that can be quantified and studied. And as a result, we will be able to determine Bayesian methods for making better decisions, given our current architecture and our preferences.

I think questions like this are slightly misguided in the following way. Humanity took a tiny step toward Bayesianism and then totally dominated the world-wide ecosystem in an evolutionary millisecond. Are we "very" Bayesian? Well, what do you mean by "very"? We certainly employ Bayesian methods more so than a hermit crab or a dolphin, but we have a long way to go if we want most of our decision making to be enacted through genuine Bayesian reasoning.

So I would say that "being Bayesian" can fall on a spectrum. Humans are not near the 0%-Bayesian end of the spectrum, but probably also not near the 100%-Bayesian perspective.

A more interesting questions is whether humanity should strive to engineer itself to specifically be more Bayesian. Some see a perfect Bayesian reasoner as a fixed point of Darwinian evolution (because the universal Solomonoff prior is not computable, it's reasonable to think such a fixed point cannot physically exist, but that doesn't mean we shouldn't try to get there anyway.)

  • $\begingroup$ Please cite comments like comparisons of humans to crabs or dolphins in terms of Bayesianism. Both of those do basic Bayesian inference for certain tasks. Dominated what? Most of the biomass on Earth is bacteria and plants. Your last paragraph is completely misguided. I cannot think of a single time when a serious biologist has talked about a 'fixed-point' in Darwinian evolution (apart from simple mathematical models). Assuming there is a fixed point, and 'striving towards it', and other forms of teleology are specifically what biologist try to avoid since it is unscientific and unfounded. $\endgroup$ Apr 6, 2012 at 23:18
  • 1
    $\begingroup$ Further, it is not clear how this answers the question about biological plausibility of bayesian models of cognition... instead of being just a general comment on Bayesianism and maybe evolution. $\endgroup$ Apr 6, 2012 at 23:21
  • $\begingroup$ You're seriously misunderstanding me and I dislike the downvote based on misunderstandings. I'm speaking of a fixed point of the computational process that is Darwinian evolution, and like I pointed out, complexity theory gives us good reason to say such a fixed point doesn't exist biologically. I wouldn't expect a biologist to be concerned with it. It's a matter for Bayesian epistemology and just a tangent to my other point. If you want to score species according to biomass, that's fine, but in terms of optimization power humans clearly win. I think that's a more relevant score. $\endgroup$
    – ely
    Apr 6, 2012 at 23:48
  • $\begingroup$ Also, my experience with this is in terms of focus of attention mechanisms, such as Bayesian surprise and the Barlow Infomax principle (which is linked to experiments done on primate cortices). I have a limited amount of additional experience in Bayesian models for primate face recognition. See here, here, here, and here for some references. None address crabs, though. $\endgroup$
    – ely
    Apr 6, 2012 at 23:54
  • 1
    $\begingroup$ My downvote is because of two reasons (1) it is not clear how the post answers the question, (2) the last paragraph is misleading. You can think of evolution however you like, but there is no reason to believe in it having fixed points, let alone bayesian inference having to do anything with them. Sure we can make models where such a fixed point exist, but we can also make selection models without fixed points, or ones where Bayesian inference will not be optimal. So if we want our models to be meaningful, we still have to defer to biologist, who in my experience would be pretty skeptical. $\endgroup$ Apr 7, 2012 at 0:07

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.