Evolutionary Neural Architecture Search for High-Dimensional Skip-Connection Structures on DenseNet Style Networks
Damien O’Neill, Bing Xue, Mengjie Zhang
Abstract
Convolutional neural networks hold state-of-the-art results for image classification, and many neural architecture search algorithms have been proposed to discover high performance convolutional neural networks. However, the use of neural architecture search for the discovery of skip-connection structures, an important element in modern convolutional neural networks, is limited within the literature. Furthermore, while many neural architecture search algorithms utilize performance estimation techniques to reduce computation time, empirical evaluations of these performance estimation techniques remain limited. This work focuses on utilizing evolutionary neural architecture search to examine the search space of networks, which follow a fundamental DenseNet structure, but have no fixed skip connections. In particular, a genetic algorithm is designed, which searches the space consisting of all networks between a standard feedforward network and the corresponding DenseNet. To design the algorithm, lower fidelity performance estimation of this class of networks is examined and presented. The final algorithm finds networks that are more accurate than DenseNets on CIFAR10 and CIFAR100, and have fewer trainable parameters. The structures found by the algorithm are examined to shed light on the importance of different types of skip-connection structures in convolutional neural networks, including the discovery of a simple skip-connection removal, which improves DenseNet performance on CIFAR10.