Hyperparameter Optimization for Convolutional Neural Networks by a Genetic Algorithm

Abstract

Convolutional Neural Networks (CNNs) achieve high discrimination performance in object recognition. Although CNNs automatically learn parameters by a data-driven method, appropriate hyperparameters must be selected to achieve high discrimination performance. However, since it is difficult to find the optimal hyperparameters by manual adjustment, studies on automatic adjustment using a search method have been actively conducted. In this research, we propose a method to optimize the number of filters in CNNs by using Genetic Algorithm (GA). Specifically, the commonly-used network architecture is used as a part of the initial population of GA for efficient search. In simulation experiments, we compare the classification accuracy of CNNs optimized by the proposed method with those by Optuna and Hyperopt.