Perceptual distinction in an unsupervised neural network: implications for theories of category-specific deficits

Abstract

There are many reports of patients who, after sustaining brain damage, exhibit a selective recognition deficit for certain categories of object. There has been much controversy as to whether this is informative about the neural organisation of knowledge in the human brain. In this paper we describe an unsupervised neural network model that is trained to process images from a variety of different object categories. Analysis of the unsupervised representations reveals some interesting distinctions between different classes of object. We contend that this model indicates a natural perceptual distinction between certain object categories, which may become exaggerated by the effects of human brain damage.