An Efficient and Flexible Automatic Search Algorithm for Convolution Network Architectures

Abstract

As many superior convolutional neural networks (CNNs) have been proposed in recent years, CNNs have played an important role in computer vision. However, manually-designing CNN architecture is difficult since expertise is required. Therefore, several automatic search algorithms have been proposed for neural architecture search, which usually have considerable computational complexity and the search space is limited. To address these problems, an efficient and flexible CNN architecture search algorithm (EF-CNN) is proposed in this paper. In EF-CNN, a flexible architecture search space is constructed by considering the depth, width, and lightweight blocks. In order to improve the reliability of the architecture while reducing the computational time, a multi-objective fitness correction method is proposed in EF-CNN based on the divided datasets, where the accuracy and computational complexity of architecture are considered simultaneously to design CNN. The experimental results on CIFAR-10 and CIFAR-100 indicate that the performance of CNN architecture designed by EF-CNN is very competitive while the computational time is greatly reduced.