On the classification capability of a dynamic threshold neural network.

Abstract

This paper proposes a new type of neural network called the Dynamic Threshold Neural Network (DTNN) which is theoretically and experimentally superior to a conventional sigmoidal multilayer neural network in classification capability. Given a training set containing 4k + 1 patterns in Rn, to successfully learn this training set, the upper bound on the number of free parameters for a DTNN is (k + 1)(n + 2) + 2(k + 1), while the upper bound for a sigmoidal network is 2k(n + 1) + (2k + 1). We also derive a learning algorithm for the DTNN in a similar way to the derivation of the backprop learning algorithm. In simulations on learning the Two-Spirals problems, our DTNN with 30 neurons in one hidden layer takes only 3200 epochs on average to successfully learn the whole training set, while the single-hidden-layer feedforward sigmoidal neural networks have never been reported to successfully learn the given training set even though more hidden neurons are used.