The Study Of The Dynamics Of Neural Networks With Fixed Degree Of Dilution

This paper extend a recently presented method of designing asymmetry neural networks with associate memory, designing a diluted neural network which has fixed degree of dilution, and study the dynamics of this diluted neural network, and compare its performance with full connected neural network.In traditional feed-back neural network, the common problem is that there are spurious memories. Then a new method was presented to design asymmetric neural network, which is called Monte Carlo-"MC" adaptation rule. And by extending this rule, designing full-connected and diluted neural network. By comparing the dynamics of the two neural network, finding that the performance of diluted neural network was poorer than the full-connected neural network. So we guess whether the performances of a diluted neural network are similar with a small scale full-connected neural network. Then we extend this MC-adaptation rule and design a diluted neural network with fixed degree of dilution using the anneal dilution method. And then analyse the dynamics of this diluted neural network. Results indicate that using this designing rule, asymmetric full connected neural network do have significant advantages over the asymmetric diluted neural network in storage ratio and study time.