Autonomous deep learning: A genetic DCNN designer for image classification

Recent years have witnessed the breakthrough success of deep convolutional neural networks (DCNNs) in image classification and other vision applications. DCNNs have distinct advantages over traditional solutions in providing a uniform feature extraction-classification framework to free users from troublesome handcrafted feature extraction. However, DCNNs are far from autonomous, since their performance relies heavily on the handcrafted architectures, which also requires a lot expertise and experience to design, and cannot be continuously improved once the tuning of hyper-parameters converges. In this paper, we propose an autonomous and continuous learning (ACL) algorithm to generate automatically a DCNN architecture for each given vision task. We first partition a DCNN into multiple stacked meta convolutional blocks and fully connected blocks, each of which may contain the operations of convolution, pooling, fully connection, batch normalization, activation and drop out, and thus convert the architecture into an integer code. Then, we use genetic evolutionary operations, including selection, mutation and crossover to evolve a population of DCNN architectures. We have evaluated this algorithm on six image classification tasks, i.e., MNIST, Fashion-MNIST, EMNIST-Letters, EMNIST-Digits, CIFAR10 and CIFAR100. Our results indicate that the proposed ACL algorithm is able to evolve the DCNN architecture continuously if more time cost is allowed and can find a suboptimal DCNN architecture, whose performance is comparable to the state of the art.