Removal of hidden units and weights for back propagation networks

The objective of this paper is to present a simple and effective method for removal of both hidden units and weights. In this paper, we propose two methods, "consuming energy" method and "weights power" method, and compare them with the conventional method. According to our computer simulations using the mirror symmetry problem, the weights power method has shown the best performance in respect of size reduction (removal of units and weights), generalization performance, and the amount of computation required. For example, the number of hidden units reduced to about 40% of the initial state, and the number of weights reduced to less than a fourth of the initial state. In addition, generalization performance was improved more than 10%.