A high-speed learning method for analog neural networks
Takazumi Matsumoto, M. Koga · 1990
A novel learning method is proposed, which is based on hardware and applicable to analog neural networks. All the network parameters (weights and thresholds) are oscillated slightly with different frequencies, and the spectra detected in the error signal are fed back to the neural network, where the network parameters are changed by a value proportional to the coherent detection signal. Since all parameters are changed in parallel, and the hardware used for learning is composed of analog circuits, this method is suited to high-speed learning. Using exclusive-OR network simulations, the effects of crosstalk between parameter control signals are discussed