Two phase learning technique in modular neural network for pattern classification of handwritten Hindi alphabets

Abstract

Listen

Translate article

Modular neural network overcomes the problem of monolithic structures of artificial neural networks. Generally modular neural network is an integration of smaller sub complete neural network models. Each model works independently on a sub portion of larger size pattern vectors. There are two ways of modularizing the neural network i.e. modularizing learning and modularizing structure. In this present work the modular neural network with modular learning for pattern classification of hand written Hindi alphabets is considered. In the presented approach 24 individual sub neural networks have been considered for first phase computing. In the second phase the collective outputs of first phase is presented as input to global neural network. Thus, the output of second phase presents the desired classification of the given large training set. Neural networks of first phase are trained locally for decomposed input patterns with gradient descent learning. Updated weights of the first phase are mapped to the global neural network. The global neural network is further trained for the collective output patterns of the first phase computing. Two phases of modular neural network i.e. decomposition and replication have been applied to perform the classification task. Simulation results are indicating that the complete neural network is approximated well when updated weights of first phase are combined with new weights of second phase and it generalized well when only updated weights of sub neural network of first phase are mapped to the connection strength of global neural network.