Dept. of Computer Science and Engineering
Oregon Graduate Institute of Science & Technology
Current interconnection architectures are not adequate to support the communications requirements of Artificial Neural Networks based upon Neurophysiological models. For ANN models, direct implementation has a cost in required area which scales as the cube of the number of connections per node. A system of one million nodes, each connected to one thousand others, would require 40 times as much silicon if implemented as a series of direct wires as it would with multiplexed interconnect. This thesis further shows that using a broadcast communication paradigm improves cost-performance results by at least a factor of N[superscript Â½] over point-to-point. Broadcast also allows for fewer messages, shorter messages, easier implementation, and can be implemented either with a physical broadcast interconnection structure or as a virtual model imposed upon a point-to-point physical interconnection architecture. This research lays the theoretical foundations for development of broadcast as an effective communications paradigm for ANN implementations. In support of the primary results of this thesis, methods of analyzing target models and interconnection architectures are developed. Other proposed interconnection architectures are compared with the proposed broadcast solutions and shown to be inadequate for these network models. In addition, results are given which show the effectiveness of broadcast for implementing ANN models ranging from artificial models such as feed-forward layered networks to olfactory piriform cortex to mammalian hippocampus to abstract neurophysiological models. It is shown that a complete rat hippocampus could be implemented in a single eight inch wafer with a .3 micron technology.
Bailey, James L., "A VLSI interconnect strategy for biologically inspired artificial neural networks" (1993). Scholar Archive. 159.