Evolution of Convolution Neural Network Architectures using Genetic Algorithm

Convolution Neural Network (CNN) has become a remarkable tool in solving many real-world problems of computer vision in recent years. However, the designing of most of CNN architectures is manual, which requires a significant trial and error methodology as CNN's performance highly relies on their architectures. Thus, it gets arduous to design a promising CNN architecture without having sufficient domain knowledge and human expertise. In this paper, we attempt to explore the possibility of using Genetic Algorithm to design CNN architectures automatically. The CNN architecture proposed by the Genetic Algorithm is trained from scratch using Gradient- Descent Algorithm and evaluated on a validation set at each evolutionary step. The algorithm does not require any preprocessing of the data before its execution, nor any post-processing on the evolved CNN architecture. We propose an encoding scheme for determining the layer connectivity. This scheme allows the formation of skip connections within the CNN architecture. Along with the encoding scheme, the filter dimensions, and the number of nodes in the fully-connected layer are also genetically evolved during the subsequent evolution using standard genetic operators, namely selection, mutation, and crossover. We have specified algorithms for the formation of the CNN model by combining the information about the encoding scheme, filter dimensions, and the number of nodes of the fully-connected layer in addition to the genetic operations.The proposed algorithm is tested on the MNIST dataset for handwritten digit recognition and Fashion-MNIST dataset. Our experiments have shown that the algorithm is capable of successfully generating high-quality CNN architectures, which are less studied before.