The subjected SPDS algorithm of multi-layer forward neural network

In the realization of the stimulant integrate circuit using multi-layer forward neural network, the problem of the network redundancy must he solved. The weights of the network are usually established by the changeable resistances, whose domain is limited, which leading to the fixed domain of the weights, that is to say, a subjected condition of the feasible domain of the weights is added to the network. While the method of deriving the weights and the layer-training algorithm has the disadvantages of the slow speed of convergence, as well as the demerit that the weights and threshold values tend to be out of the feasible domain. in this paper, the subjected SPDS algorithm based on the idea of circulating the coordinate in turns is presented, which makes sure that the network parameters He in the feasible domain for all the time, furthermore, in relatively short time, the training requirement can be realized and it is good for hardware design of neural network.