VLSI implementation of neural networks with application to signal processing

Mapping a functional neural network model to analog sub-threshold MOS technology is a challenging task, and requires careful architectural, system level and circuit level consideration, with respect to the constraints inherent in this technology. The authors present their experience in this mapping process. The artificial neural network systems addressed are programmable ones facilitating learning either on or off chip. The authors consider multi-layer feedforward networks, although the techniques can be easily adapted to recurrent networks. A multi-layer learning algorithm suitable for analog sub-threshold implementation is presented. The authors discuss system level issues, describe circuits of neurons and synapses that have been designed, and present fabrication results