ABSTRACT
One productive and influential approach to cognition maintains that categorization, object recognition, and higher-level cognitive processes operate on the output of lower-level perceptual processing. Computational models of cognition can often be thought of as consisting of representation-process pairs. And it seems almost axiomatic that one needs to determine or “fix” representations if this enterprise is to succeed. Indeed cognitive scientists are often quite skilled at creating situations where conjectures about representations receive support and attention can focus on processing principles. Models of concepts generally assume that concepts are represented in memory in terms of components that are parts of a code used to describe them. The chapter discusses computational issues in the learning of new features of representation, as well as connectionist mechanisms that integrate feature discovery and other aspects of cognition. An active area of connectionist research involves the investigation of unsupervisedlearning algorithms that discover regularities in their environment.
