Feed-forward Neural Network Classifiers with Bithreshold-like Activations

Abstract

The paper deals with the issues concerning the modifications of the bithreshold neuron whose activation functions provide better ability of the solving the classification problems. The model of smoothed local bithreshold neuron is proposed, which is capable to recognize compact finite set of patterns in n-dimensional space. We design a binary classifier on the base of the feed-forward neural network whose hidden layer consists of such neurons with modified activations, propose the synthesis algorithm and estimate its time complexity alongside with the networks size. The simulation results demonstrate that the application of modified activations improves the accuracy of classification.