A Neural Implementation of Optimal Cue Integration

This chapter lays out a theoretical framework for how optimal cue integration can be implemented by neural populations. The main significance of this framework does not merely lie in understanding multisensory perception in a principled manner, but in the fact that it provides a blueprint for finding neural implementations of other forms of Bayes-optimal computation. Evidence for Bayesian optimality of human behavior has been found in many perceptual tasks, including decision making, visual search, oddity detection, and multiple-trajectory tracking. Probabilistic population coding provides a roadmap for identifying a neural implementation of each of these computations: First the Bayesian model at the behavioral level needs to be worked out, then it needs to be assumed that probability distributions in this model are encoded in neural populations with Poisson-like variability, and finally the neural operations that map onto the desired operations on probability distributions should be identified.