Neural Architecture Search with Improved Genetic Algorithm for Image Classification

Abstract

Neural Architecture Search (NAS) is an automatic process of designing a neural architecture for solving classification problems. It is closely related to hyper-parameters such as hidden layers, neurons in each hidden layer, type of activation function (ACT), network optimizer and so on. Therefore, finding appropriate hyper-parameters to construct suitable network architecture for a particular problem is a challenging task. In this paper, an improved Genetic Algorithm (GA-NAS) is proposed to build a multi-layer feed forward architecture for image classification problem. Each chromosome of the proposed method is encoded with four hyper-parameters namely no. of hidden layers, neurons per hidden layer, activation function (ACT) and network error optimization technique. Each chromosome represents a neural network architecture for the given problem. The categorical cross-entropy or log function is considered to represent fitness function which provides performance accuracy of the architecture. The proposed methodology is experimented on two well-known benchmark image classification data sets such as CIFAR-10 and MNIST. The GA-NAS is compared with brute force algorithm and obtained results demonstrated the effectiveness for solving image classification problems.