PHD-NAS: Preserving helpful data to promote Neural Architecture Search

Shun Lu, Yu Hu, Longxing Yang, Jilin Mei, Zihao Sun, Jianchao Tan, Chengru Song

Abstract: Neural Architecture Search (NAS) has achieved promising results in many domains. However, the enormous computational burden consumed by the NAS procedure significantly hinders its application. Existing works focus on mitigating the search cost by either designing a more efficient algorithm or searching in an elaborately designed search space, heavily relying on expert experience and domain knowledge. We notice that few works focus on dataset optimization for NAS, however, the truth is that not all samples are essential for the search process, which can be omitted actually. Therefore, we propose to only preserve helpful data for the supernet training to improve the efficiency. Specifically, we compute the forgetting and remembering events for each sample during the supernet training to determine the data importance. Samples that the supernet has predicted correctly in consecutive epochs have low importance and will be gradually removed from the dataset during training. We further formulate our method into a unified cycled-learning framework for jointly optimizing proxy dataset and architecture search. By combining with different algorithms, we demonstrate that our framework can find architectures with comparable performance using much less training data and search time in various search spaces and benchmarks, validating the effectiveness of our method.