Estimation of initial weights and hidden units for fast learning of multilayer neural networks for pattern classification

A method is proposed for weight initialization in backpropagation feedforward networks. Training data is analyzed and the notion of critical point is introduced for determining the initial weights and the number of hidden units. The proposed method has been applied to artificial data and the publicly available cancer database. The experimental results of artificial data show that the proposed method takes 1/3 of the training time required for standard backpropagation. In order to verify the effectiveness of the proposed method standard backpropagation, where the learning starts with random initial weights, was also applied to the cancer database. The experimental results indicate that the proposed weight initialization method results in better generalization