Electronic Grey Matter

Abstract

One of the most exciting long term goals of Electrical Engineering is to construct an electronic machine with functional ability similar to that of the human brain. A brief inspection of the brain indicates the enormous difficulties involved. The brain is overlaid with a thick covering or cortex of grey matter comprising approximately 1010 neurons or active elements plus additional regulating circuitry in the cerebellum. The electrical engineer setting out to duplicate such a gadget is faced with three stark problems: (1) Wired in cost per active element. The cost of 1010 at a penny each (wired in) would be $100 million for hardware. (2) Power dissipation per active element. The power dissipation of 1010 neurons at 1 μW each would be 10 kW. (3) Design scheme. Some plan of construction is obviously necessary before one begins to build something with 1010 working parts. If the design requires every interconnection to be correctly made, it will be impossible to carry it through. The main purpose of this paper is to suggest that there is now the possibility of overcoming these difficulties using tunnel junctions between superconducting metals as the active elements.1,2 These devices are analyzed in detail and it is demonstrated both analytically and experimentally that neuristor pulse propagation is possible. The junctions have extremely low power dissipation per unit area (about 400 μW/cm2 in the resting state for normal electron tunneling and zero dissipation if the resting state for Josephson or superelectron tunneling) which solves problem 2. The use of modern photoresist fabrication techniques should satisfy problem 1. Finally, problem 3 has been discussed in detail by Beurle3 who has assumed a mass of randomly interconnected active trigger circuits with a learning capability (provided, for example, by an excitation threshold which decreases with use). He has shown that large scale wave and oscillatory behavior occur which can be arranged to provide both conditioned response and memory patterns of time sequences of events. In this paper we also describe a saturable ferromagnetic interconnection which is compatible with the superconducting neuristors and which will provide the necessary learning capability required for the construction of the systems considered by Beurle.