The Subjected SPDS Algorithm of Forward Neural Network

The network redundancy is the key problem in the realization of the stimulant integrate circuit using multi-layer forward neural network. It is well known that the weights of the neural network are usually established by the changeable resistances, whose limited range leads to the fixed domain of the weights. This means that a subjected condition of the feasible domain of the weights is added to the neural network. The method of deriving the weights and the layer-training algorithm does well in training neural network, but it has the disadvantages of the slow speed of convergence, as well as the weights and threshold values tend to be out of the feasible domain. Deriving the weights and the layer-training algorithm seems to be good, but it can´t always keep the weights in the feable domain. The subjected SPDS algorithm based on the idea of circulating the coordinate in turns works well and it assures all the network parameters keep in the feasible domain for all the time. Furthermore, the training requirement can be realized in relatively short time, and it is fit for hardware design of neural network. The thesis makes a deep analysis of the subjected SPDS algorithm, and presents the method to improve its training speed,and numerical experimentation tests its validity.