Neuroscience ↔ Statistics

The brain implements approximate Bayesian inference — perception equals likelihood times prior divided by evidence — and neural populations encode probability distributions, making predictive processing (Helmholtz's unconscious inference) a formal instantiation of Bayes' theorem in cortical circuits.

ESTABLISHED
neuroscience statistics cognitive-science Bayesian-inference computational-neuroscience

🔭 Overview

Helmholtz (1867) proposed that perception is "unconscious inference" — the brain uses prior knowledge to resolve ambiguous sensory input. This informal insight has been formalised into the Bayesian brain hypothesis: sensory perception is the computation of the posterior P(cause | sensation) ∝ P(sensation | cause) × P(cause) where P(cause) is the prior expectation and P(sensation | cause) is the likelihood of the sensory input given a hypothesised cause. FORMAL CORRESPONDENCES: 1. PERCEPTION = BAYESIAN INFERENCE: Weiss et al. (2002) showed that the perceived direction and speed of moving stimuli (including classic visual illusions) are exactly predicted by a Bayesian observer with a prior favouring slow motions. The "slow motion prior" is the neural prior P(speed = slow). Psychophysical data fits Bayesian optimal observers with surprising precision across modalities (vision, touch, audition, multisensory integration). 2. DIVISIVE NORMALIZATION = BAYESIAN NORMALIZATION: The divisive normalization operation ubiquitous in cortical neurons — response = input / (gain + Σ inputs) — is the neural implementation of computing a normalized probability distribution. Carandini & Heeger (2012) showed divisive normalization is the canonical neural computation; Bayesian normalization (dividing by the evidence P(sensation)) explains why the brain uses this seemingly arbitrary nonlinearity. 3. ATTENTION = PRECISION WEIGHTING: Friston's Free Energy Principle formalises attention as the precision weighting of prediction errors — the inverse variance (precision = 1/σ²) of the sensory likelihood. When attention is directed to a stimulus, its precision weight is increased, amplifying the prediction error signal that drives perceptual updating. This maps directly onto the Fisher information of the sensory channel. 4. PREDICTION ERRORS = SURPRISE SIGNALS: The difference between predicted and actual sensory input (residual = observation - prediction) is the Bayesian prediction error. In the brain, this is computed by superficial pyramidal cells in cortical layer II/III, which project bottom-up; predictions are carried by deep layers projecting top-down. The laminar structure of cortex implements the message-passing algorithm for approximate Bayesian inference. 5. HALLUCINATIONS AND PRIORS: Psychotic hallucinations arise when the prior dominates over the likelihood — the brain "hallucinates" a percept that matches its prior regardless of incoming sensory evidence. Predictive processing predicts this: increased prior precision (relative to likelihood) causes perception to be driven by top-down priors alone. This generates a testable pharmacological prediction: drugs that reduce likelihood precision (e.g. sensory deprivation or ketamine, which blocks NMDA receptors that carry prediction errors) should increase hallucination rates. 6. NEURAL POPULATION CODES AS PROBABILITY DISTRIBUTIONS: Knill & Pouget (2004) proposed that neural population activity represents a probability distribution — the posterior over the underlying variable. The mean firing rate encodes the mean of the posterior; the width of the population activity pattern encodes posterior variance. The Cramér-Rao bound constrains how precisely any neural code can represent a parameter given its Fisher information.

⚙️ The Mathematical Bridge

This bridge connects Neuroscience and Statistics through shared mathematical structure. Status: Established connection.

↔️ Translation Table

Domain A Term Domain B Term Note
Prior P(cause)Neural prior — baseline expectation from past experience and top-down contextEncoded in the synaptic weights of top-down connections in cortical hierarchies
Likelihood P(sensation | cause)Sensory evidence carried by bottom-up afferent signalsPrecision-weighted prediction errors in layer II/III superficial pyramidal cells
Posterior P(cause | sensation)Percept — the brain's inferred cause of the sensory inputApproximately computed by loopy belief propagation in cortical circuits
Bayesian updating (prior → posterior)Perceptual learning — prior is updated by prediction errors over timeCorresponds to Hebbian-like synaptic changes driven by persistent prediction error
Precision (inverse variance 1/σ²)Attention — relative gain of prediction error signalsHigh-precision likelihood = high attention to sensory channel; modulated by ACh and NE neuromodulators
Prediction error (residual)Surprise signal — ascending from lower to higher cortical areasMinimizing prediction error = perception + action; Free Energy Principle of Friston (2005)
Bayesian model selection (evidence)Perceptual hypothesis testing — why does the brain choose one interpretation over another?Model evidence = log marginal likelihood; the brain minimises free energy ≈ surprisal

🗺️ Why Hasn't This Been Unified?

Helmholtz's insight (1867) predated Bayesian statistics; the formal Bayesian brain formulation arrived with Knill & Pouget (2004) and Friston (2005) but remains specialised literature in computational neuroscience. Most experimental neuroscientists record single neurons rather than measuring population-level probabilistic codes; the statistical framework requires population recordings and Bayesian model fitting that most labs lack. The free energy principle literature (Friston) is mathematically dense and has a reputation for being difficult to interpret, slowing adoption by the experimental community.

🌱 Cross-Pollination Opportunities

Open Questions

📚 References