Bandit-NAS: Bandit Sampling Method for Neural Architecture Search

Existing NAS (Neural Architecture Search) algorithms achieve a low error rate on vision tasks such as image classification by training each child network with equal resources during the search. However, it is not necessary to train with the equal resource or use the fully converge score to obtain the relative performance of each child network, and there is computational redundancy in training all child networks with the equal resource. In this paper, we propose Bandit-NAS to automatically compute the required data slicing and training time for each child network. i): We first model the search of the best child network training time for a given resource into an M-armed bandit problem. ii): Then we propose a reward-flexible bandit algorithm in conjunction with existing reinforcement learning-based NAS algorithms to determine an update strategy. The proposed Bandit-NAS can train M child networks simultaneously under a given resource constraint (training time for one epoch), and the amount of training data is allocated according to the current accuracy of the child networks, thus minimizing the error rate of the child networks. Experiments on CIFAR-10 show that proposed Bandit-NAS performs better the baseline NAS algorithm, e.g., ENAS, with lower error rate and faster searching time.