REPRESENTATION OF VARIABLES IN CONNECTIONIST NETWORKS.

There are various types of fine-grain parallel architectures which are proving useful for artificial intelligence research, ranging from neural networks, through connectionist networks, to the Connection Machine. The style of computing that is possible with all such connectionist networks is different from that possible with uniprocessor systems, or with parallel systems containing a small number of processors. Just as a different style of computation is possible in fine-grain parallel systems, the styles of representation of variables that are natural for such parallelism are different from the types of representation natural for serial, or coarse-grain parallel processing. However, there has been no general theoretical analysis of connectionist representations. A language is provided in which different parallel representations can be accurately described and the differences between representations clarifed. The theoretical framework for variable representations is also useful for suggesting new types of representations. A theoretical framework is developed which can aid in the formal analysis of fine-grain parallel representations.