The Parallel Distributed Processing Approach to Modeling Cognition and its Neural
 Basis

James L. McClelland, Stanford, USA

This workshop will provide an overview of the Parallel Distributed Processing approach to modeling cognition, including applications to cognitive development and effects of brain damage and disease on cognitive processes.  I will present the PDPtool software environment, with live simulations of processing and learning in PDP networks.  I will relate the PDP framework to other levels of analysis (detailed neural modeling, rational analysis) and emphasize how the models provide alternatives to other perspectives on the nature of processing, learning, and representation.

The workshop will have four parts.  In part 1, the idea that cognitive processes arise from bi-directional interactions among simple processing units will be discussed, evidence that the brain adheres to this principle will be presented, and the state of recent debates on this topic will be reviewed.  In part 2, principles of connection-based learning and distributed representation will be reviewed, with emphasis on recent developments including the unsupervised learning of internal representations in deep belief networks. In part 3, the application of these ideas in language will be considered.    It will be argued that language itself and human language processing both exhibit a wide range of gradient phenomena, motivating connectionist alternatives to symbol- and rule-based architectures.  In part 4, the PDP framework will be applied to development of conceptual knowledge in children and to disintegration of conceptual knowledge in semantic dementia. Relationships between connectionist and structured probabilistic approaches to knowledge representation, and applications to the role of metaphor in language and thought will be explored in this context.