The Bayesian brain hypothesis, developed by David Knill and Alexandre Pouget (2004) and building on Helmholtz's (1867) notion of "unconscious inference," proposes that the brain represents and processes sensory information according to the principles of Bayesian probability theory. On this view, perception is not a passive registration of stimuli but an active inference process: the brain combines prior expectations P(s) with incoming sensory evidence P(x|s) to compute a posterior belief P(s|x) about the most probable state of the world.
Bayesian Inference in Perception
Bayesian estimate: ŝ = arg max P(s|x) (MAP)
or ŝ = ∫ s · P(s|x) ds (posterior mean)
Optimal integration of two cues:
ŝ = (σ₂² · s₁ + σ₁² · s₂) / (σ₁² + σ₂²)
σ_combined² = (σ₁² · σ₂²) / (σ₁² + σ₂²)
A landmark prediction of the Bayesian framework is optimal cue combination: when two sensory cues provide information about the same quantity, the optimal estimate is a reliability-weighted average, with weights proportional to the inverse variance (precision) of each cue. Ernst and Banks (2002) confirmed this prediction in a visual-haptic size estimation task, showing that humans weight visual and haptic information according to their respective reliabilities, shifting from vision-dominated to haptic-dominated estimates as visual noise is increased.
Priors and Perception
Bayesian models explain a wide range of perceptual phenomena as consequences of prior expectations. The "light-from-above" prior biases the interpretation of shading patterns, causing concavities to be perceived as convexities when shadows are inconsistent with overhead illumination. The "slow motion" prior (Weiss, Simoncelli, & Adelson, 2002) explains why low-contrast stimuli appear to move more slowly: when the sensory likelihood is broad (uncertain), the posterior is pulled toward the prior expectation of slow speed.
The Bayesian brain framework subsumes classical Signal Detection Theory as a special case. In SDT, the optimal observer computes the likelihood ratio and compares it to a criterion determined by prior probabilities and payoffs — precisely the Bayesian posterior odds. The Bayesian framework extends SDT by allowing for structured priors, hierarchical inference, and the integration of information across time and modalities, providing a more general account of perceptual decision making.
Neural Implementation
A central challenge for the Bayesian brain hypothesis is specifying how neural circuits implement Bayesian computations. Proposed neural coding schemes include probabilistic population codes (Ma, Beck, Latham, & Pouget, 2006), in which the pattern of activity across a neural population implicitly represents a probability distribution, and sampling-based representations, in which neural variability reflects samples from the posterior distribution. Both frameworks make testable predictions about the relationship between neural variability and perceptual uncertainty.
Critics of the Bayesian brain hypothesis note that human behavior frequently deviates from Bayesian optimality — people are overconfident, exhibit base-rate neglect, and show systematic biases. However, proponents argue that these deviations can often be explained by approximate inference algorithms (sampling, variational methods) operating under computational constraints, or by misspecified priors. The framework's power lies not in claiming that humans are optimal but in providing a normative benchmark against which deviations can be precisely characterized and explained.