A model of categorization and learning in a connectionist broadcast system

A computational model of some aspects of categorization and learning is developed for a connectionist network, i.e., a network of simple computing elements among which memory and control are distributed. The network is called the broadcast net. The questions addressed in this dissertation are important to cognitive modeling: How can a connectionist network represent arbitrary temporary associations? How are fast novel perceptions possible in a network that learns slowly? How can the network maintain several temporary associations without crosstalk among them? Equivalently, how can it maintain the pairings between several "roles" and their "role-fillers" without confusing which filler is bound to which role? In the broadcast net, temporary associations of features are represented in a different coding from that in which durable information is represented. This allows novel patterns of activation to be established quickly and to temporarily coexist with previously learned patterns in the network. The broadcast net explicitly represents the distinction between the association of features in a single pattern, and the juxtaposition of different simultaneous active patterns. The broadcast net is a pattern-completion network that completes and maintains several distinct patterns simultaneously. This allows it to maintain several arbitrary temporary associations without crosstalk, and therefore to maintain the pairings between roles and role-fillers. Learning in the broadcast net is the recoding of information from its representation by temporary associations of features to a durable representation by weights on connections among computing elements.