Examination of implementing a neural network on a parallel computer — NEOCOGNITRON on NCUBE
Takayuki Itō, Kunihiko Fukushima, Sei Miyake · Systems and Computers in Japan · 1991
Abstract The parallel computer based on the parallel operation of a large number of CPU is now available as a means of realizing a high‐speed computer. Especially, the loosely coupled parallel computer is considered to be interesting as a tool for realizing the neural network model, since it is related closely to the idea of the group of neurons. In such a computer, the efficiency is affected greatly by the assignment of functions to CPU. From such a viewpoint, this paper discusses the realization of a high‐speed neural network model on a loosely coupled parallel computer. The neocognitron is considered as the neural network model; and the structure of the model, together with the feasibility of parallel operation, is discussed. By utilizing the feature of the loosely coupled parallel computer with hypercubic connection, an efficient CPU assignment is considered. Based on those results, a recognition system for handwritten numerals by the neocognitron is realized on the parallel computer NCUBE, and the recognition speed of eight characters/s is realized. Finally, the data transfer method and the load balancing among CPU are discussed, aiming at the realization of an exhaustive connection, which is required, in general, in the neural network model.