Comparison of Harris Performance as Activation Function to Rectified Linear Unit (ReLU), Leaky ReLU, and Tanh in Convolutional Neural Network for Image Classification

Abstract

Activation functions (AFs) are the building blocks of a deep neural network (DNN) to perform image classification effectively by handling nonlinear data and extracting complex features and patterns. This paper introduces a new activation function (AF) called “Harris” a piecewise and nonlinear nonmonotonic AF inspired from the field of photonics. The AF was integrated to a simple convolutional neural network (CNN) using Canadian Institute for Advanced Research (CIFAR-10) dataset to determine the model’s performance in terms of accuracy in training and testing, image classification capability, and feature extraction.  Harris was able to exceed the leaky Rectified Linear Unit (ReLU) and hyperbolic tangent function (tanh) target accuracies from α-values −0.80 to −1.00 in image classification, while the testing accuracies were able to exceed the target accuracies of ReLU from α-values −0.80 to −0.95. It was able to handle negative values solving the dead neuron problem and extract complex features through its feature maps which improve the F1-scores of the CNN model in image classification.