Deeper Insights into Weight Sharing in Neural Architecture Search

TL;DR

This paper investigates the effects of weight-sharing in neural architecture search, revealing its impact on variance, interference, and performance, and suggesting ways to mitigate issues for better model evaluation.

Contribution

It provides a comprehensive experimental analysis of weight-sharing in NAS, highlighting its drawbacks and proposing strategies to reduce variance and improve results.

Findings

High variance in model performance due to weight-sharing

Interference among child models causes high variance

Reducing weight-sharing degree improves stability and performance

Abstract

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.