Collective computation in neuronlike circuits.

M odern digital computers are latecomers to the world of computation. Biological com­ puters-the brain and nervous sys­ tem of animals and human beings­ have existed for millions of years, and they are marvelously effective in processing sensory information and controlling the interactions of ani­ mals with their environment. Tasks such as reaching for a sandwich, recognizing a face or remembering things associated with the taste of madeleines are computations just as much as multiplication and running video games are. The fact that biological computa­ tion is so effective suggests that it may be possible to attain similar ca­ pabilities in artificial devices based on the design principles of neural systems. We have studied a number of "neural network" electronic cir­ cuits that can carry out significant computations. Such simple models have only a metaphorical resem­ blance to nature's computers, but they offer an elegant, different way of thinking about machine computa­ tion, which is inspiring new micro­ electronic chip and computer de­ signs. They may also provide fresh insights into the biological systems. Current research on this subject builds on a long history of efforts to capture the principles of biological computation in mathematical mod­ els. The effort began with the pio­ neering investigations of neurons as logical devices by Warren S. McCul­ loch and Walter H. Pitts in 1943. In the 1960's Frank Rosenblatt of Cornell University and Bernard Widrow, who is now at Stanford University, creat­ ed "adaptive neurons" and simple networks that learn. Widrow's Ada­ line (short for adaptive linear ele-