MMI-based-training for a probabilistic neural network
Nan Bu, Toshio Tsuji, Osamu Fukuda · 2004
Probabilistic neural networks (PNNs) that incorporate the Bayesian decision rule and statistical models have been widely used for pattern classification. Efficient estimation of the PNN's weights, however, is still a major problem. In this paper, we propose a new training scheme based on a discriminative criterion, maximum mutual information (MMI), and apply this method to the log-linearized Gaussian mixture network (LLGMN) which is one of the PNNs. The MMI training achieves a consistent estimator of network weights, and includes the conventional maximum likelihood (ML) algorithm as a special case. Also, the dynamics of terminal attractor (TA) is introduced for iteration control of the MMI training. Finally, the classification ability of the proposed method is examined with a pattern classification problem of the electromyogram (EMG) signals, and found that the MMI training results in better classification than the conventional ML algorithm.