Improving Performance of Deep Learning Models using Input Propagation

Abstract

Artificial Neural Networks (ANNs) act by fitting decision boundaries between different classes of data, thereby performing classification. They approximate the decision hyperplane into a weighted sum of inputs, thereby making the decision simpler. This paper explores the working of ANNs in a simple scenario where the decision boundary is quadratic, finds that weighted sum won't be enough to perform non-linear classifications since decision boundaries remain linear in nature. To overcome the limitations of the existing model, an Input-Propagated Artificial Neural Network (IP-ANN) is proposed, which learns nonlinear functions directly by incrementally applying input values on the network directly. The performance of the proposed model is compared with the regular ANNs, and experimental analysis shows that the regular ANN has obtained an accuracy of 72.36%, while IP-ANN achieved an accuracy of 98.5%.