Dr. Michael Lewicki-Colloquium Announcement

Learning abstract representations of structures in natural scenes

 

Michael Lewicki

Electrical Engineering & Computer Science

Case Western Reserve University

 

We readily perceive contours and surfaces in complex natural scenes.  At the level of simple visual features, however, these more abstract structures are difficult to extract, because image patterns of both boundaries and surface regions are highly variable.  What then are the computations that can deduce intrinsic structure from the raw sensory variability?  In this talk, I will discuss an approach that is based on learning statistical distributions of both local regions in a visual scene and its global structure.  This approach generalizes the theory of efficient coding for learning image features to hierarchical models.  The central hypothesis is learning these local distributions allows the visual system to generalize across similar local image regions, i.e. textures within a surface or texture boundaries along a contour.  Joint activity in the model encodes the probability distribution over their inputs and forms stable representations across complex patterns of variation.   In addition, units in the model exhibit a diverse range of non-linear properties observed in complex cells in visual cortex and offer a novel functional explanation for their role in visual perception.

 

Dr. Lewicki works jointly with Yan Karklin, Chris DiMattina, and Wooyoung Lee.