Papex

Open problem · Neuroscience · Open

Bayesian brain

Does the brain make sense of the world by making predictions according to the rules of Bayesian probability?

Summary

In some perceptual and motor tasks the nervous system appears to combine prior expectations with uncertain sensory evidence close to the statistical optimum. Whether the brain really computes in this way, and how neurons would do it, is open.

Why it is open

According to the Bayesian brain hypothesis, neural circuits combine prior knowledge with new evidence, weigh sources of information by their reliability and take actions that account for uncertainty [13]. A noisy or ambiguous measurement gives a wide likelihood function in Bayesian terms, and prior knowledge is used to resolve the ambiguity where possible [13]. David Knill and Alexandre Pouget called the neural version the "Bayesian coding hypothesis": "that the brain represents sensory information probabilistically, in the form of probability distributions" [2].

The problem has two levels. Behaviour in several perceptual and motor tasks is close to what a statistically optimal observer would do [1, 3]. Whether neurons compute this way is a separate question: in 2004 Knill and Pouget called neurophysiological data on the hypothesis "almost non-existent" [2], and in 2012 Jeffrey Bowers and Colin Davis saw impressive mathematical analyses of how neural populations could compute in a Bayesian manner "but little or no evidence that they do" [7]. Even the behavioural level is disputed: Dobromir Rahnev and Rachel Denison review suboptimal performance in eight classes of perceptual decisions [10].

The hypothesis is also stated in different ways. Richard Lange and colleagues distinguish theories that ask how circuits could recover information about the world from neural activity (Bayesian decoding) from theories that ask how circuits implement inference in an internal model of the world (Bayesian encoding); leaving the difference unstated, they argue, makes general conclusions difficult [13].

No body sets a bar for a resolution. Edgar Walker and colleagues propose four criteria for a neural representation of uncertainty: activity should change with uncertainty (sensitivity), still do so when confounding features such as image contrast are held fixed (specificity), not change with unrelated features (invariance), and cause behaviour (functionality) [14]. Rahnev and Denison argue that the field should stop asking whether behaviour is optimal and build observer models that explain behaviour across many tasks [10].

What is known

Behaviour (2002–2004). Marc Ernst and Martin Banks measured how variable people's estimates of an object's height were by sight alone and by touch alone, built from these a maximum-likelihood integrator, a rule that weights each cue so that the combined estimate varies least, and found that it "behaved very similarly to humans" when both senses were used [1]. Konrad Körding and Daniel Wolpert controlled the statistical distribution of a new sensorimotor task and the uncertainty of the sensory feedback; their subjects represented both and combined them "in a manner consistent with a performance-optimizing bayesian process" [3]. Cue-combination studies across vision, touch, hearing, the vestibular sense and proprioception have been "fundamental to the view that sensory perception is optimal", Rahnev and Denison write [10].

Neural proposals (2006–2011). Wei Ji Ma, Jeffrey Beck, Peter Latham and Pouget argued that the variability of cortical neurons means that populations "automatically represent probability distributions over the stimulus", a scheme they called probabilistic population codes, and showed that with Poisson-like variability a broad class of Bayesian inference reduces to linear combinations of population activity [4]. In sampling models, neural activity over time traces out samples from the represented distribution rather than encoding its parameters [13]. Pietro Berkes, Gergő Orbán, Máté Lengyel and József Fiser related evoked activity to inferences and spontaneous activity to prior expectations in an internal model, and predicted that the two should match if the model is statistically optimal; in the visual cortex of awake ferrets the similarity increased with age and was specific to natural scenes [6]. Karl Friston's free-energy principle, proposed to account for action, perception and learning, treats surprise, or prediction error, as the quantity the brain optimises, and suggests that several brain theories might be unified within it [5]. Laurence Aitchison and Lengyel describe the combination of predictive coding and Bayesian inference as "Bayesian predictive coding" [9]; the cortical evidence for predictive coding is covered on the page on canonical cortical computation.

Neural tests (2020–2025). Walker, James Cotton, Ma and Andreas Tolias recorded from primary visual cortex (V1) of two macaque monkeys with 96-electrode arrays during an orientation categorisation task in which trial-by-trial uncertainty mattered [12]. The likelihood function decoded from population activity on each trial predicted decisions better than a point estimate of orientation, also when the true orientation was held fixed [12]. In 2025 Charles Findling, Félix Hubert and the International Brain Laboratory analysed brain-wide Neuropixels recordings and widefield calcium imaging in mice that reported a visual stimulus appearing on the left or right with a prior probability alternating between 0.2 and 0.8 in blocks [15]. The mice estimated this prior, and it was encoded in at least 20% to 30% of brain regions, from early sensory areas to motor and high-level cortical regions; the authors read this as consistent with Bayesian inference involving loops between areas, rather than a prior added only in decision-making areas [15]. How neural activity carries information in general is the problem of the neural code.

Critiques (2012–2018). Bowers and Davis argued that priors, likelihoods and utility functions can be altered in many arbitrary ways to fit the data, making the models unfalsifiable, and that Bayesian theories rarely predict data better than simpler non-Bayesian ones [7]. Thomas Griffiths, Nick Chater, Dennis Norris and Pouget replied that the critique did not accurately characterise the motivation behind Bayesian modelling and did not evaluate it against specific alternatives [8]. Rahnev and Denison's suboptimalities include misweighting in cue combination: in one auditory–visual localisation study, for example, observers treated the visual cue as though it were more reliable than it was [10]. They also note statements from prominent Bayesian theorists that their goal is not to demonstrate optimality [10].

Main approaches

Probabilistic population codes. The population that encodes a stimulus also carries the uncertainty about it, so that a likelihood function can be read from its activity on each trial [4, 12]. Strongest evidence: likelihoods decoded from macaque V1 that predict decisions beyond a point estimate [12]. Main difficulty: such tests are "code-driven": they make assumptions about the neural code for the world state and its uncertainty, while correlational approaches impose no such constraint, and the two lead to different but complementary findings [14]. Lange and colleagues place these codes in the decoding perspective, which treats neural activity as given and asks what an observer of it could infer [13].

Sampling. Activity over time represents samples from the probability distribution over the causes of sensory input in an internal model, and spontaneous activity reflects the prior [6, 13]. Strongest evidence: the developmental convergence of spontaneous and natural-scene-evoked activity in ferret visual cortex [6]. Main difficulty: to predict neural data, such models need a further linking hypothesis on how samples map onto membrane potentials and spikes [13]; and the debate between sampling and parametric codes is complicated because sampling codes usually make encoding assumptions while parametric codes often make decoding ones [13].

Predictive coding and free energy. Perception, action and learning optimise one quantity, surprise or prediction error [5], and predictive coding is one proposed way to implement Bayesian inference [9]. Strongest evidence: Aitchison and Lengyel call predictive coding an idea that "accounts for a remarkable range of phenomena" [9]. Main difficulty: they judge the evidence "inconclusive in the sense that it does not rule out Bayesian inference with a direct variable code", and argue that predictive coding can serve goals other than Bayesian inference [9]. Samuel Gershman finds that the free-energy principle "does not have a fixed set of distinctive claims" and makes different claims under different assumptions, which must be checked in each application for its claims to be falsifiable [11].

Models without optimality. Behaviour, including its departures from optimality, is modelled without assuming that the brain is optimal, by non-Bayesian theories [7] or by detailed observer models [10]. Strongest evidence: suboptimalities reported "in virtually every class of perceptual tasks" [10]. Main difficulty: Griffiths and colleagues reply that the critique of Bowers and Davis misdescribed the aims of Bayesian modelling and did not test it against specific alternatives [8].

Open sub-questions

  1. Likelihoods beyond V1. Does the trial-by-trial likelihood decoded from populations in higher visual or decision areas predict choices better than a point estimate, as Walker et al. found in macaque V1 [12]?
  2. Where prior meets likelihood. In the International Brain Laboratory task, do the regions that encode the block prior [15] also carry trial-by-trial sensory uncertainty, as local Bayesian combination would require?
  3. A functional prior. Does silencing a region that encodes the prior reduce the mice's use of it, meeting Walker et al.'s functionality criterion [14, 15]?
  4. Spontaneous activity and image statistics. Does the match between spontaneous and natural-scene-evoked activity found in developing ferrets [6] weaken in animals reared with altered image statistics?
  5. Misweighting in neurons. When observers overweight a cue relative to its measured reliability [10], does the cue's neural representation show the higher reliability that behaviour implies?
  6. Separating predictive coding. Which experiment would distinguish Bayesian predictive coding from Bayesian inference with a direct variable code, which Aitchison and Lengyel say the current evidence does not rule out [9]?

Work on this in Papex

If you have a testable answer to one of the sub-questions above, you can draft it as a private hypothesis in Papex, or prepare the draft from Claude or ChatGPT with the Papex connector. You review the exact version, sign it and decide whether to publish it; a public record that names this problem is listed on this page. A Papex record shows who signed which version and when. It is not peer review and does not mean the hypothesis is correct.

Key references

  1. Ernst, M. O. & Banks, M. S. (2002). Humans integrate visual and haptic information in a statistically optimal fashion. Nature 415(6870), 429–433. 10.1038/415429a · Checked 15 Oct 2026
  2. 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. 10.1016/j.tins.2004.10.007 · Checked 15 Oct 2026
  3. Körding, K. P. & Wolpert, D. M. (2004). Bayesian integration in sensorimotor learning. Nature 427(6971), 244–247. 10.1038/nature02169 · Checked 15 Oct 2026
  4. Ma, W. J., Beck, J. M., Latham, P. E. & Pouget, A. (2006). Bayesian inference with probabilistic population codes. Nature Neuroscience 9(11), 1432–1438. 10.1038/nn1790 · Checked 15 Oct 2026
  5. Friston, K. (2010). The free-energy principle: a unified brain theory? Nature Reviews Neuroscience 11(2), 127–138. 10.1038/nrn2787 · Checked 15 Oct 2026
  6. Berkes, P., Orbán, G., Lengyel, M. & Fiser, J. (2011). Spontaneous cortical activity reveals hallmarks of an optimal internal model of the environment. Science 331(6013), 83–87. 10.1126/science.1195870 · Checked 15 Oct 2026
  7. Bowers, J. S. & Davis, C. J. (2012). Bayesian just-so stories in psychology and neuroscience. Psychological Bulletin 138(3), 389–414. 10.1037/a0026450 · Checked 15 Oct 2026
  8. Griffiths, T. L., Chater, N., Norris, D. & Pouget, A. (2012). How the Bayesians got their beliefs (and what those beliefs actually are): comment on Bowers and Davis (2012). Psychological Bulletin 138(3), 415–422. 10.1037/a0026884 · Checked 15 Oct 2026
  9. Aitchison, L. & Lengyel, M. (2017). With or without you: predictive coding and Bayesian inference in the brain. Current Opinion in Neurobiology 46, 219–227. 10.1016/j.conb.2017.08.010 · Checked 15 Oct 2026
  10. Rahnev, D. & Denison, R. N. (2018). Suboptimality in perceptual decision making. Behavioral and Brain Sciences 41, e223 (published online 2018). 10.1017/S0140525X18000936 · Checked 15 Oct 2026
  11. Gershman, S. J. (2019). What does the free energy principle tell us about the brain? Neurons, Behavior, Data analysis, and Theory 2(3). 10.51628/001c.10839 · Checked 15 Oct 2026
  12. Walker, E. Y., Cotton, R. J., Ma, W. J. & Tolias, A. S. (2020). A neural basis of probabilistic computation in visual cortex. Nature Neuroscience 23(1), 122–129 (published online 2019). 10.1038/s41593-019-0554-5 · Checked 15 Oct 2026
  13. Lange, R. D., Shivkumar, S., Chattoraj, A. & Haefner, R. M. (2023). Bayesian encoding and decoding as distinct perspectives on neural coding. Nature Neuroscience 26(12), 2063–2072. 10.1038/s41593-023-01458-6 · Checked 15 Oct 2026
  14. Walker, E. Y., Pohl, S., Denison, R. N., Barack, D. L., Lee, J., Block, N., Ma, W. J. & Meyniel, F. (2023). Studying the neural representations of uncertainty. Nature Neuroscience 26(11), 1857–1867. 10.1038/s41593-023-01444-y · Checked 15 Oct 2026
  15. Findling, C., Hubert, F., International Brain Laboratory, Acerbi, L., Benson, B., Benson, J. et al. (2025). Brain-wide representations of prior information in mouse decision-making. Nature 645(8079), 192–200. 10.1038/s41586-025-09226-1 · Checked 15 Oct 2026

Page reviewed 15 Oct 2026

Source

Open, as listed by Wikipedia · Unsolved problems in neuroscience. Checked 08 Oct 2026.

The status is the source’s. Papex does not decide whether a problem is solved.

Records on this problem

No public record addresses this problem yet.

More in Life sciences