On-chip learning in neurocomputers

Artificial neural networks (ANNs) may be implemented as custom analog, digital or hybrid VLSI systems. This paper describes the tradeoffs among these approaches, based on work in our laboratory as well as at other institutions. A major theme of the work is the effects of limited precision in on-chip learning computations performed by the analog or digital circuits. Analog and low-precision digital circuits are found to be capable of reliably representing most ANN models, with area-efficient and energy-efficient implementations