Abstract
In this paper, we propose a new network element named a “distance-based cosine element” for neural networks. We also derive a learning algorithm based on the back-propagation algorithms for multilayer networks. The distance-based cosine element inputs a squared distance between an input pattern vector and its weight vector, and uses an affine transformation of cosine function as its output function. The proposed distance-based cosine network is able to improve its learning speed as well as convergence rate because its output function does not have any saturated regions which cause slow learning speed of the back-propagation learning using sigmoid elements.
We demonstrate the advantages of the proposed network by solving N-bits parity problems and Fisher's Iris classification problem. Experimental results indicate that our distance-based cosine network consistently obtains better results than the conventional sigmoid network in terms of both the learning speed and the convergence rate.
