Head of the Computational Systems Neuroscience Lab at the Department of Computational Sciences, MTA Wigner Research Center for Physics in Budapest, Hungary
Wednesday 20 december at 11:00 am
ISCMJ amphi
Top-down effects in visual cortical computations
Top down influences are ubiquitous in the sensory system, yet their functional role is poorly understood. In particular, the most successful theoretical framework, deep-discriminative models, lack a theoretical grounding of how top-down influences might contribute to an efficient computational machinery for processing natural images. Here we consider an alternative framework, deep generative models and capitalize on recent advances in machine learning to develop a hierarchical model of V1 and V2 computations of the dorsal stream. We point out a normative role of top-down connections when performing hierarchical inference and discuss its computational consequences. We discuss the development of a nonlinear representation in a natural-image trained deep generative model. Beyond the characterization of mean responses, we explore the role of response variance, noise correlations, and spontaneous activity in this framework. Finally, we discuss, the interplay between these perceptual top-down computations, and attention-induced top-down effects in V1.
