Weight initialization with reference patterns

Abstract

The problem of weight initialization in multilayer perceptron networks is considered. A new computationally simple weight initialization method based on the usage of reference patterns is presented. A reference pattern is a vector which is used to represent data points that fall in its vicinity in the data space. On one hand, the proposed method aims to set the initial weight values to be such that inputs to network nodes are within the active region (in other words, nodes are not saturated). On the other hand, the goal is to distribute the discriminant functions formed by the hidden units evenly into the input space area where training data is located. The proposed method is tested with the widely used two-spirals classification benchmark problem and channel equalization problem where several alternatives for obtaining suitable reference patterns are investigated. Also, the effect of the initialization is studied when two commonly used cost functions are used in the training phase. These are the mean square error and relative entropy cost functions. A comparison with the conventional random initialization shows that significant improvement in convergence can be achieved with the proposed method. In addition, the computational cost of the initialization was found to be negligible compared with the cost of training.