Favoring Generalists over Specialists: How Attentional Biasing Improves Perceptual Category Learning

Abstract

A model of cortical learning is proposed, which incorporates supervised feedback using two forms of attention: (i) feature-specific attention which allows the network to learn associations between specific feature conjunctions (or categories) and outputs, and (ii) nonspecific attentional "vigilance" which biases this learning when the associations appear to be incorrect. Attentional vigilance improves learning if it favors, via modulatory weights, generalist categories over specialist categories. A biologically plausible neural network is proposed which implements these computational principles and which outperforms several classifiers on classification benchmarks.