Abstract
Behavioral and electrophysiology studies of shape processing have demonstrated greater sensitivity to differences in non-accidental properties (NAPs) than metric properties (MPs; see Biederman, 2007 for review). NAPs correspond to image properties that are invariant to changes in out-of-plane rotation (e.g., straight vs. curved contours) and are distinguished from metric properties (MPs) that can change continuously with variations over depth orientation (e.g., aspect ratio, degree of curvature, etc). Previous work has shown that such sensitivity is incompatible with hierarchical models of object recognition such as HMAX (Riesenhuber & Poggio, 1999; Serre et al, 2007), which assume that shape processing is based on broadly tuned neuronal populations with distributed symmetric bell-shaped tuning: Shape-tuned units in these models are modulated at least as much by differences in MPs as in NAPs (Amir, Biederman & Hayworth, 2012). Here we test the hypothesis that simple mechanisms for learning transformation sequences may increase sensitivity to differences in NAPs vs. MPs in HMAX. We created a database of video sequences of objects rotated in depth in an attempt to mimic sequences viewed during object manipulation by infants during early developmental stages. We adapted a version of slow feature analysis (Wiskott & Sejnowski, 2002) to learning in HMAX: Unit responses in intermediate processing stages were scaled according to how stable they remained during the presentation of common objects undergoing various transformations. We show that this simple learning rule leads to shape tuning in higher stages with greater sensitivity to differences in NAPs vs. MPs consistent with monkey IT data (Kayaert et al, 2003). Overall we propose a simple learning mechanism to extend hierarchical models of object recognition to exhibit greater sensitivity for NAPs than MPs, as observed both behaviorally and electrophysiologically. Our results suggest that greater sensitivity for NAPs may result from unsupervised learning mechanisms from transformation sequences of common objects.
Meeting abstract presented at VSS 2014