A general-purpose neural network with on-chip BP learning
Chun Lu, Bingxue Shi, Chen Lü · 2003
A general purpose expandable neural network chip with on-chip BP (back-propagation) learning is designed. The unit chip has 4 neurons and 16 synapses. A large-scale neural network with arbitrary layers and discretional neurons per layer can be constructed by combining many unit chips. A novel neuron circuit with programmable parameters is proposed. It generates not only the sigmoid function but also its derivative. The neuron has a push-pull output stage to gain strong driving ability in both charge and discharge processes, which is very important in heavy load situations. The unit chip is fabricated with a standard 0.5-/spl mu/m CMOS, double-poly, double-metal technology. The learning system itself can be used as a refresh tool to keep the weight value right. The results of parity experiments show that it can accomplish on-chip BP learning.