The Research And Application Of The Capacity Of Hopfield Associative Memory Neural Networks

Artificial neural network is a mathematical model that would simulates the structure of thebrain synapses connection and processes related information. As a kind of neural networks,Hopfield associative memory neural networks can recover the correct and complete informationfrom the complete,defaced, noised information which the traditional method is difficult to dowith，taking advantage of massively parallel processing and good tolerance. However, thetraditional Hopfield associative memory neural network is not satisfying when it is applied to thepattern recognition of incomplete and noised samples the low memory capacity. The work of thispaper is closely around above problems and the paper’s main research results are as follows:1) When Hopfield network is used in pattern recognition of incomplete and aberrantsamples in traditional way, the correct rate of recognition is low. To solve this problem, animprovement method for storage capacity of Hopfield network based on clonal selectionalgorithm is put forward. Firstly, the clonal selection algorithm is introduced into Hopfieldnetwork, and the initial input samples of network are used as antigen of clonal selectionalgorithm. Secondly, weight matrix which is generated at random is used as the initial antibodyof the clonal selection algorithm. Thirdly, cloning, crossover and mutation are operated on theinitial antibody and the weights are optimized by using affinity. Finally, this method is applied inpattern recognition of incomplete and aberrant samples.2) The proposed method was applied to the pattern recognition of incomplete and noisedsamples to check the effectiveness. The experimental results show that in the conditions ofwithout changing the network structure, learning rules and capacity definition, the networkreferred in this paper could acquire more memory capacity which did not need to multiplemathematical iterative and two kinds of evolutionary fitness to act on under the same conditions.3) When the discrete Hopfield network is used in pattern recognition of incomplete andnoised letter, the correct rate of recognition is low. The paper proposed an optimize idea toimprove the network algorithm with immune genetic algorithm, and presents a new Hopfieldnetwork model based on immune genetic algorithm.4) The method is applied in the pattern recognition of26English letters. Experimentalresults show that the correct memory ability of Hopfield network based on hybrid immunealgorithm put forward in this paper accounts for more than90%.85%even in the case that thenoise_lever is up to0.6,0.8or1. The experiment results show its method is much effective.