Challenges in Understanding Visual Shape Perception and Representation: Bridging Subsymbolic and Symbolic Coding

Abstract

Perceiving and representing the shapes of contours and objects are among the most crucial tasks for biological and artificial vision systems. Much is known about early cortical encoding of visual information, and at a more abstract level, experimental data and computational models have revealed great deal about contour, object, and shape perception. Between the early “subsymbolic” encodings and higher level “symbolic” descriptions (e.g., of contours or shapes), however, lies a considerable gap. In this chapter, we highlight the issue of attaining symbolic codes from subsymbolic ones in considering two crucial problems of shape. We describe (1) the dependence of shape perception and representation on segmentation and grouping processes. We show that in ordinary perception, shape descriptions are given to objects rather than visible regions, and we review progress in understanding interpolation processes that construct unified objects across gaps in the input. We relate these efforts to neurally plausible models of interpolation, but note that current versions still lack ways of achieving symbolic codes. We then consider (2) properties that (some) shape representations must have and why these require computations beyond the local information obtained in early visual encoding. As an example of how to bridge the gap between the subsymbolic and symbolic, we describe psychophysical and modeling work in which contour shape is approximated in terms of constant curvature segments. Our “arclet” model takes local, oriented units as inputs and produces outputs that are symbolic contour tokens with constant curvature parts. The approach provides a plausible account of aspects of contour shape perception, and more generally, it illustrates the kinds of properties needed for models that connect early visual filtering to ecologically useful outputs in the perception and representation of shape.