Angela Huang¹, Patrick Garrigan², Philip Kellman¹; ¹University of California, Los Angeles, ²Saint Joseph's University

Recent work on contour interpolation in visual completion distinguishes a bottom-up contour-linking process—producing an intermediate representation based on edge and junction geometry—and a subsequent, more probabilistic, constraint satisfaction–based scene description process that operates on this representation to resolve border ownership, closure, and crossing interpolations (Kellman & Fuchser, 2023). Models of the contour-linking process based on neurally plausible filters have been proposed (Heitger et al., 1998; Kalar et al., 2010). However, these models produce pictorial activation maps—discrete points of edge, keypoint, and interpolation activity. Observers viewing these outputs see connected edges, illusory contours, closure, and shape, but the models themselves lack any symbolic tokens corresponding to these entities. In contrast, symbolic descriptions appear necessary for implementing subsequent scene constraints. Here we present a new model of the transition from subsymbolic filter outputs to a symbolic representation of the linked contour contour representation. We employed the biologically plausible front end of Heitger et al. (1992) to generate initial activation maps and applied simple mathematical operations to extract precise edge locations and orientations, as well as keypoint locations (contour junctions). Interpolation is implemented using abstract curvature elements hypothesized to encode smooth contours (Baker, Garrigan & Kellman, 2021). The model thus embodies a concrete proposal about the transition from subsymbolic to symbolic representations. Empirical tests show improved agreement with human perception. Whereas prior neurally-based completion models used only positive cases—stimuli in which human observers perceive illusory contours—we tested both positive and negative cases. The results indicate that the new model closely matches human judgments of the presence or absence of illusory contours. In contrast, models lacking a transition to symbolic coding generate substantial interpolation activity both for displays in which human observers see illusory contours and those in which they do not.