E15 Inference and Reconstruction of the World

Topic

This perception does not constitute a passive copy of the environment but rather an estimate constructed from sensory information and prior knowledge. According to a Bayesian interpretation, the nervous system combines both sources to infer the most probable state of the environment. This principle helps explain illusions, perceptual constancies, and context effects, especially when sensory information is ambiguous. If prior knowledge is more precise or carries greater relative weight than new sensory evidence, the final estimate leans more heavily toward that prior expectation.

Combining different sensory sources can improve the precision of this estimate. When two measurements provide partially independent information about the same variable, integrating them—weighted according to their uncertainty—can yield an estimate with lower variance than either measurement alone. This principle helps explain the advantages of binocular vision, binaural hearing, and the combination of vision, the vestibular system, and proprioception during postural control. If one source is more precise, it exerts greater influence on the resulting perception in an approximately optimal integration.

Meanwhile, a generative model represents the sensory signals that would be produced by various possible environmental states. In predictive processing frameworks, internally generated predictions are continuously compared with the signals actually received, generating prediction errors when the two differ. These errors can be used to update internal estimates and reduce future discrepancies. A larger difference between a prediction and the sensory information—depending on the precision attributed to each—increases the signal available to correct the internal model.

The Kalman filter provides a mathematical example of this combination of prediction and observation for linear systems subject to Gaussian noise. Its gain determines the weight assigned to a new measurement relative to the previous prediction, based on the uncertainty associated with each. If the measurement is precise, the estimate is strongly adjusted toward it; if the measurement is noisy, the prediction based on the prior state carries more weight. Some models of perception and motor control employ similar principles to describe how the nervous system estimates the state of the body and the environment based on imperfect signals.

These mechanisms allow for the interpretation of various perceptual phenomena. In perceptual constancy, the brain uses contextual information to estimate relatively stable object properties despite changes in conditions such as lighting. In multisensory integration, a more reliable modality may be weighted more heavily and partially dominate the combined perception. In the rubber hand illusion, sufficiently coherent visual, tactile, and proprioceptive signals can temporarily alter the brain's estimate of body position and ownership. Greater spatial and temporal coherence among these multisensory signals makes it more likely that the perceptual system will attribute them to a common cause and, consequently, update its body representation.

ID:3901