Deeper Insights into Weight Sharing in Neural Architecture Search
This work addresses the lack of theoretical guarantees for weight-sharing in NAS, an incremental improvement for researchers and practitioners in automated machine learning.
The paper investigates the impact of weight-sharing in Neural Architecture Search, revealing that it causes high variance in model performance but can still provide useful information, and shows that reducing weight-sharing reduces variance and improves performance.
With the success of deep neural networks, Neural Architecture Search (NAS) as a way of automatic model design has attracted wide attention. As training every child model from scratch is very time-consuming, recent works leverage weight-sharing to speed up the model evaluation procedure. These approaches greatly reduce computation by maintaining a single copy of weights on the super-net and share the weights among every child model. However, weight-sharing has no theoretical guarantee and its impact has not been well studied before. In this paper, we conduct comprehensive experiments to reveal the impact of weight-sharing: (1) The best-performing models from different runs or even from consecutive epochs within the same run have significant variance; (2) Even with high variance, we can extract valuable information from training the super-net with shared weights; (3) The interference between child models is a main factor that induces high variance; (4) Properly reducing the degree of weight sharing could effectively reduce variance and improve performance.