Optimum associative neural network utilizing maximum likelihood
Hong Chang,Xiaoliang Wei,Zuren Feng
DOI: https://doi.org/10.1016/j.neucom.2008.02.001
IF: 6
2009-01-01
Neurocomputing
Abstract:In this paper, a new learning rule and theoretical analysis of an extended bidirectional associative memory network (MLBAM) is presented, by using the maximum likelihood criterion based on two well recognized and essential criteria, i.e., the convergence of the learning rule, and the noise tolerance of this network. Traditional methods fail to distinguish highly approximative patterns. However, the method in our study can improve this by using the newly developed method of maximum likelihood criterion. To be specific, by employing the MLBAM, association and memory could be clearly distinguished. In addition, the learning approach guarantees that correlated patterns could be associated as a stable state and the network possesses excellent anti-noise property by using likelihood function, namely, the learning approach specializes in the situation including stochastic disturbance. Additionally, the associative capability of the bidirectional associative memory is specifically discussed. Finally, three experiments are used to certify the validity and efficiency of our method, especially the method's excellent anti-noise property by using likelihood function.