Evolutionary Designing CNN Architectures Using The Thermodynamical Genetic Algorithm

Abstract

Convolutional Neural Network (CNN) is one of useful methods for image recognition. For the task of setting hyperparameters becoming huge and complex with the development of CNN, the evolutionary acquisition approach to CNN architecture using Genetic Algorithm (GA), a method to efficiently optimize the problem based on evolutionary dynamics, has attracted much attention. In this study, we propose Thermodynamical Genetic Algorithmic Convolutional Neural Networks (tdgaCNN), which is a CNN architectural search method based on Thermodynamic Genetic Algorithm (TDGA), one of the methods for GA. This method achieves better CNN architecture acquisition by emphasizing the importance of maintaining population diversity during the GA search phase. The effectiveness of the proposed method is confirmed by an image benchmark dataset. We also apply the proposed method to classification tasks among trompe l'oeil as typical images giving multiple views, landscapes and portraits.