Abstract

In recent years, a growing number of human psychophysical studies and physiological experiments with behaving animals supported the notion that the brain encodes both the value and the uncertainty of the stimulus during perception. As a result, cortical sensory processing has been increasingly viewed in the framework of probabilistic inference. However, there have been much fewer attempts to extend this probabilistic notion to learning even though perception and learning are clearly two fundamental aspects of cortical processing. In addition, present theoretical probabilistic models of cortical inference cannot handle easily important aspects of cortical activity such as trial-to-trial variability, and structured spontaneous activity. I will present a framework and some results that address these issues and potentially link learning behavior to probabilistic neural activity in the cortex. First, I provide evidence that human learning of chunks, new visual feature combinations is close to statistically optimal. Next, I present a generative framework of cortical processing and learning based on the assumption that neural activity represents samples of the posterior probability distribution defined by the uncertainty of the stimulus. This framework can explain trial-to-trial variability, gives a functional role to spontaneous activity in the cortex, and provides some clearly testable physiological predictions. In the final part of my talk, I will demonstrate how we confirmed one of these predictions by showing that with accumulating visual experience, the distribution of spontaneous activity in the cortex approximates the distribution of evoked activity averaged over natural stimuli.