We are studying how shape information is represented by large populations of cells at intermediate and higher stages in the ventral pathway of primate visual cortex. Our results show that shapes are represented in terms of their component parts. A parts-based coding scheme has the combinatorial power to represent an infinity of shapes with a finite number of neurons (just as 26 letters of the alphabet represent thousands of words). Many ventral pathway neurons are tuned for parts-level information about boundary curvature, medial axis conformation, and relative position of shape compo-nents. A given cell, for example, might respond to all shapes that contain downward-pointing convex projections near the right. these tuning properties can be characterized with mathematical functions in high-dimensional shape space, which allows us to reconstruct the population representation of any given shape by combining signals from large populations of cells. In this way we can read out the neural code for object shape from the response rates of neurons.

Faculty Host: Dr. Barbara Landau (landau@cogsci.jhu.edu)