Perception as an Inference Problem

Although the idea of thinking of perception as in inference problem goes back to Helmholtz, it is only recently that we have seen the emergence of neural models of perception that embrace this idea. Here I describe why inferential computations are necessary for perception, and how they go beyond traditional computational approaches based on deductive processes such as feature detection and classification. Neural models of perceptual inference rely heavily upon recurrent computation in which information propagates both within and between levels of representation in a bi- directional manner. The inferential framework shifts us away from thinking of 'receptive fields' and 'tuning' of individual neurons, and instead toward how populations of neurons interact via horizontal and top-down feedback connections to perform collective computations.