Dept. of Computer Science and Engineering
Oregon Graduate Institute of Science & Technology
Artificial neural networks constitute attempts to put computer systems on the road to natural intelligence. Promising applications await even limited progress in this direction, as challenging problems in perception, control, and reasoning resist traditional computational approaches. Unfortunately, the capabilities and organizations of today's artificial neural networks are far removed from those of their biological namesakes. This difference will narrow, however, as increasing understanding of nervous system operation is incorporated into artificial neural network design. The hardware performance required by such networks will not be met by conventional computers; new architectures will be needed that support the processing requirements of nervous system organization. An investigation is described of neurocomputer architectures and circuit designs required to implement a neural network modeled directly from studies in mammalian cortex. Previous research in electronic neural network hardware has targeted neurons performing peripheral sensory roles, such as in the auditory and visual systems, or has viewed neural network operation as routine (albeit massively parallel) vector processing. The cortical model used here is the piriform model of olfactory/piriform cortex, developed by Richard Granger and Gary Lynch at the University of California, Irvine. The "piriform model" is actually a spectrum of models ranging from biochemically detailed descriptions to abstract signal processing versions divorced from biology. Two points in this range will be studied. The first is a high-level model description that has been shown to perform a useful signal processing function known as vector quantization. An implementation of this model on a commercial neurocomputer will be described. The second version of the model, containing more biological detail, is the primary focus of the architectural and circuit design effort. A neurocomputer architecture called "Super Sniff" (SS) is proposed, and two possible implementations are studied. One features analog processing and direct connections between neurons, and the other utilizes a shared communication structure and digital processing. The cost/performance tradeoffs of each are compared. Sparse temporal activity and connectivity, limited precision arithmetic, and a discrete-event style of operation are characteristics of the model that are compatible with VLSI processing. Mammalian olfactory cortex contains many millions of neurons; this thesis investigates networks containing more modest neuron counts numbering in the tens of thousands. Lessons learned about the suitability of VLSI processes for implementing biologically-faithful cortical structures are discussed in the conclusion.
Means, Eric, "Designs for a cortically-inspired neurocomputer architecture" (1991). Scholar Archive. 161.