EvoDCNN: An evolutionary deep convolutional neural network for image classification

Abstract

Developing Deep Convolutional Neural Networks (DCNNs) for image classification is a complicated task that needs considerable effort and knowledge. By employing an evolutionary computation approach, one can automatically generate the network models. However, the Neuroevolution is computationally expensive, and in some cases it needs hundreds of GPU days for training. Therefore, there is a need to find optimum Neuroevolutionary models with minimum computation to deal with this problem. In this paper, by utilising a Genetic Algorithm (GA), we introduce EvoDCNN, as a block-based evolutionary model for developing an evolutionary deep convolutional network for image classification. Such that by using the proposed fixed-length encoding model, we can generate variable-length networks with high accuracy while using less computation. The proposed model by utilising a straightforward evolutionary framework is able to establish small networks with high classification accuracy. Eight datasets: CIFAR10, MNIST, and six versions of EMNIST, that include balanced and unbalanced datasets, are used for evaluation of the proposed model. We did a comprehensive evaluation where we compared the results with many previous works, and outperformed the previous state-of-the-art accuracy for classification of five of the datasets.