Reading english test with spatially-invariant recptive fields: when the 'binding problem' isn't

摘要

Accumulating experimental evidence suggests that 91) recognition of complex objects and scenes can sometimes occur in a single feedforward pass through the visual system (Fize et al., 1998), and (2) the visual field may be encoded during this first pass in terms of a population of spatially-invariant detectors of visual mini-patterns (Kobatake and Tanaka, 1994). We study here some of the computational properties of visual representations based on spatially-invariant receptive fields (RF's), using the domain of text as a convenient surrogate for visual recognition in general. We begin by developing an analytical model that makes explicit how recognition performance is affected by (1) the number of object categories (workds) that must be distinguished, (2) the complexity of individual objects (length of workds), (3) the number and order of binding of elemental features (letters) included in the representation, and (4) the clutter load, i.e. the amount of visual material (text) in the field of view in which multiple objects must be recognized with-out explicit segmentation. We show that that the model achieves good fits to recognition rates for English text over a wide range of clutter loads, word sizes, and feature counts. We then show, using a quasi-supervised greedy algorithm for feature learning, that fewer than 1,500 mostly low-order, spatially-invariant letter-tuple detectors (akin to Wichelfeatures) are needed to unambiguously represent all the workds simultaneously present in randomly chosen windown tof text up to 50 characters in width. Our results help explain how, and under what conditions, spatially-invariant RF-based representations can process multiple objects simultaneously witout explicit segmentation processes, and they lend support to the notion that representations of this simple kind may underlie important aspects of primate/human recognition.