利用最大似然准则的双向联想网络研究

Bidirectional Associative Memory Utilizing Maximum Likelihood

针对现有双向联想网络(BAM)存在的存储容量小、抗干扰能力弱的缺点,提出了一种利用最大似然准则的BAM网络(MLBAM)及其训练算法.MLBAM网络采用双向网络结构建立了神经元的发放以及抑制模型,充分利用似然函数的特性以及网络的双向联想特性,很好地完成了自联想和异联想功能,并且准确计算出关联样本对之间的关联度,使MLBAM网络在随机环境中具有很强的抗噪能力.利用最速下降算法,给出了MLBAM网络的训练算法,根据训练权重的Hessian矩阵负定,判定算法能够获得全局最优解,从而证明了算法的收敛性.该训练算法能够训练出最优的连接权重和神经元阈值.通过2个典型实验验证了MLBAM网络的抗噪能力和联想能力,在存在1位随机噪声的情况下,该网络的联想正确率达到了100%.

A new learning rule and theoretical analysis of an extended bidirectional associative memory network （MLBAM） are 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 the network. Traditional methods fail to distinguish highly closing patterns. However, this disadvantage is improved by using the newly developed method, since the maxi- mum likelihood method is used to seek the best possible closing mapping of two patterns. Experiments are made to verify the validity and efficiency of the proposed method. The method displays excellent anti-noise property and MLBAM network exhibits association at a 100% accuracy under one-bit inversion which implies that 100% of the one-bit errors is corrected.