Purchase this article with an account.
Ruth Rosenholtz; The power of pooling in high dimensions. Journal of Vision 2013;13(9):627. doi: 10.1167/13.9.627.
Download citation file:
© 2016 Association for Research in Vision and Ophthalmology.
Evidence suggests that crowding results from "forced texture processing," involving excessive feature integration or compulsory averaging over each local pooling region. A number of experiments have tested a simple version of this hypothesis. In this simple pooling model, each pooling region yields the mean of some (often unspecified) feature. To a first approximation, this predicts worse performance the more one fills the pooling region with irrelevant flankers. For a given amount of "flanker stuff", performance should not depend on the perceptual organization of the flankers, nor on the flanker identities, except insofar as they affect the informativeness of the mean feature value. This impoverished model cannot explain performance improvements with larger flankers (Levi & Carney 2009; Manassi et al 2012), or when flankers group with one another (Saarela et al. 2009; Sayim et al 2010; Manassi et al. 2012). It cannot predict worse performance with increasing target-flanker similarity (Andriessen & Bouma 1976; Kooi et al 1994; Saarela et al. 2009), unless one hypothesizes that averaging occurs only within a narrow feature band. And why do observers not merely report the mean? These results seem to bode ill for pooling models of crowding.
However, the model being rejected is a straw man. Realistic texture processing models are high-dimensional, and bear little resemblance to their low-dimensional brethren. High-dimensional models behave fundamentally differently, and intuitions do not simply "scale up" from low-dimensional models. More measurements per patch provide increasingly good representation of that patch. I will demonstrate simple techniques for gaining intuitions about high-dimensional models. Such models can capture the "objectness" and numerosity of display items, aspects of their feature distributions (not just their mean), differences between target and flankers, and flanker grouping. Such information can facilitate decision-making about the target in ways not predicted by a "simple pooling model".
Meeting abstract presented at VSS 2013
This PDF is available to Subscribers Only