An algorithm to find the optimized network structure in an incremental learning

Abstract

In this paper1 we show a new learning algorithm for pattern classification. A scheme to find a solution to the problem of incremental learning algorithm is proposed when the structure becomes too complex by noise patterns included in the learning data set. Our approach for this problem uses a pruning method which terminates the learning process with a predefined criterion. Then an iterative model with a 3 layer feedforward structure is derived from the incremental model by appropriate manipulation. Note that this network is not fully connected between the upper and lower layers. To verify the effectiveness of the pruning method, the network is retrained by EBP. We test this algorithm by comparing the number of nodes in the network with the system performance, and the system is shown to be effective.