David Rumelhart

Parallel distributed processing models represent knowledge as patterns of activity spread across many simple processing units. Rumelhart and James McClelland presented this connectionist approach as an alternative to cognition based only on explicit symbolic rules.

Rumelhart, Geoffrey Hinton, and Ronald Williams published an influential account of training multilayer neural networks with backpropagation in 1986. Their work helped revive interest in learning internal representations through error-driven adjustment.

The model learned patterns linking present-tense verbs with past-tense forms without being given a conventional list of grammar rules. Its successes and errors fueled a major debate between connectionist researchers and critics such as Steven Pinker and Alan Prince.

Schemas are organized knowledge structures that guide comprehension, interpretation, and memory. Rumelhart used the concept to explain how prior knowledge helps people understand stories and fill in information that a text leaves unstated.