マージン最大化による多層ニューラルネットの学習方式

抄録

In this paper, we propose the training method for three layer neural network classifiers in which we calculate the weights by maximizing margins in each layer. According to the CARVE algorithm, we need to find a hidden layer hyperplane that separates a set of data of one class from the remaining training data. Then, the separated set of data is removed from the training data. We repeat this procedure until the remaining data belong to one class. In the proposed method we use a heuristic algorithm to train the support vector machine so that the data on one side of the hidden layer hyperplane belong to one class. For the output layer, we use a quadratic optimization technique. The performance of this new algorithm is evaluated using some benchmark data sets.