A fast GA for automatically evolving CNN architectures
Hong Zhou, Wei Hua Fang, Jun Jie Sun, Xiao‐Jun Wu · 2020
Convolutional neural networks (CNNs) have been proved to be effective models to solve a series of challenging computer vision tasks. However, designing CNN architectures with good performance is still a challenging task. Automatic CNN architecture search algorithms have been proposed in recent years which can find competitive CNN architectures without manual intervention. But automatic search algorithms usually consume considerable computational time and resources. In addition, they only use deep blocks and ignore wide blocks of CNNs, which limits the performance of evolved CNNs. In order to address the above issues, a fast approach for automatically evolving CNN architectures based on deep and wide blocks (FAE-CNN) is proposed through genetic algorithms in this paper. In FAE-CNN, a refined fitness evaluation method based on divided datasets is designed with the purpose to speed up the running time. By introducing the Inception Block, FAE-CNN can evolve optimal CNN architectures from both deep and wide directions. Experimental results on CIFAR10 and CIFAR100 show that FAE-CNN can automatically design CNNs with flexible architectures and better performance in a very short running time.