Electronic Neural Networks: A New Class of Computing Hardware

Over the last decade the semiconductor industry has seen a tremendous increase in the complexity and power of integrated circuits. These advances have been driven by the need for ever more powerful and flexible computing and made possible by steady improvements in line rules and circuit design. Since the early 70's, line rules have decreased from about 10 microns to slightly below 1 micron while circuits have advanced from 1 Kbit DRAM to 4 or even 16 Mbit DRAM today at a cost per device that has decreased over a thousand-fold. There are, however, eventual limits to this increase in complexity based on economics (the need for new technologies below about 1/4 micron line rules) or physics (quantum mechanics and discrete particle effects below about 0.1 microns [l]). Even extending technology to these extremes, though, there are still problems like speech recognition or computer vision that appear to be beyond our computational abilities. These tend to have in common a need for fast, approximate pattern matching or association within a large database, a task that is poorly suited to conventional digital hardware and architecture.