A high-dimensional pooling model accounts for seemingly conflicting substitution effects in crowding

Abstract

Visual crowding is a phenomenon in which the perception of a peripheral target depends on nearby, task-irrelevant flankers. The exact nature of this dependency is controversial (Whitney & Levi 2011). Phenomenologically, some flanker features appear to be confused, or substituted, with target features. One current debate (akin to early- vs. late-selection debates in attention) centers around whether substitution occurs for entire objects, for basic features, or something in between. Results have been mixed: in letter stimuli, observers sometimes appear to swap the target with entire flankers (van den Berg et al. 2012), to preferably swap with similar flankers, or to report letters not present in the display (Freeman et al. 2012). In feature conjunction stimuli, observers appear to swap unbound features (Põder & Wagemans 2007), whereas in stimuli of intermediate complexity observers appear to lose location information after binding simple features (Greenwood et al. 2012). We ask whether a high-dimensional pooling model can predict these seemingly disparate results. In this model (Balas et al 2009), the visual system measures a texture descriptor (Portilla & Simoncelli 2000) in each local pooling region. We test whether the ambiguities and confusions inherent in this representation can account for diverse substitution phenomena. To do this, we synthesize new images ("mongrels") which share the same texture descriptor as the original stimuli. We find that the model explains seemingly contradictory results. In mongrels of "complex" letter stimuli, we observe swapping of flankers with targets, preferred generation of similar flankers, and synthesis of absent letters. Mongrels of simple feature conjunction stimuli show illusory conjunctions, supporting observations of non-binding of features. Mongrels of arrays of crosses (Greenwood et al. 2012), typically reproduce arrays of crosses, in line with apparent low-level "binding" in crowding of these stimuli. Thus, our high-dimensional pooling model can reconcile disparate crowding accounts. Meeting abstract presented at VSS 2013