Improving generalization ability of HMM/NNs based classifiers

Standard Hidden Markov Models (HMM) have proved to be a very useful tool for temporal sequence pattern recognition, although they present a poor discriminative power. On the contrary Neural Networks (NNs) have been recognized as powerful tools for classification task, but they are less efficient to model temporal variation than HMM. In order to get the advantages of both HMMs and NNs, different hybrid structures have been proposed. In this paper we suggest a HMM/NN hybrid where the NN classify from HMM scores. As NN we have used a committee of networks. As networks of the committee we have used a Multilayer Perceptron (MLP: a global classifier) and Radial Basis Function (RBF: a local classifier) nets which drawn conceptually different interclass borders. The combining algorithm is the TopNSeg scoring method which sum the top N ranked networks normalized outputs for each class. The test of above architecture with speech recognition, handwritten numeral classification, and signature verification problems show that this architecture works significantly better than the isolated networks.