[3J4-OS-3b] AutoML（自動機械学習）(2/2)

[3J4-OS-3b-01] Stopping criterion for Neural Architecture Search

〇Kotaro Sakamoto Sakamoto¹, Hideaki Ishibashi², Rei Sato³, Shinichi Shirakawa⁴, Yohei Akimoto^3,5, Hideitsu Hino^1,5 (1. The Institute of Statistical Mathematics, 2. Kyushu Institute of Technology, 3. University of Tsukuba, 4. Yokohama National University, 5. RIKEN Center for Advanced Intelligence Project (AIP))

Keywords:Neural Architecture Search, Deep Learning

Neural architecture search (NAS) is a framework for automating the design process of a neural network structure. While the recent one-shot approaches have reduced the search cost, there still exists an inherent trade-off between cost and performance. It is important to appropriately stop the search and further minimise the high cost of NAS. On the other hand, heuristic early-stopping strategies have been proposed to overcome the well-known performance degradation of the one-shot approach, particularly differentiable architecture search (DARTS). In this paper, we propose a more versatile and principled early-stopping criterion on the basis of the evaluation of a gap between expectation values of generalisation errors of the previous and current search steps with respect to the architecture parameters. The stopping threshold is automatically determined at each search epoch without cost. In numerical experiments, we demonstrate the effectiveness of the proposed method. We stop the one-shot NAS algorithms such as ASNG-NAS and DARTS and evaluate the acquired architectures on the benchmark datasets: NAS-Bench-201 and NATS-Bench. Our algorithm has been shown to reduce the cost of the search process while maintaining a high performance.