A Novel Method for Suitable Hyperparameter Selection in an Accurate Convolutional Neural Network Architecture

Abstract

The deep convolutional neural network (CNN) models are of great use in many areas and applications such as image processing and computer vision. The hyperparameter optimization in the CNN architectures is essential for an efficient implementation of model on software or hardware or “software-hardware co-design” platform to achieve better characteristics. In this paper, we have proposed CNN architecture models trained using MNIST dataset that explores the selection of various hyperparameters and their impact on the accuracy to achieve the hyperparameter optimization. The work presents thorough evaluation of various hyperparameters which offers a higher accuracy and keeps the architecture simple as compared with other published results.