Understanding Architectures Learnt by Cell-based Neural Architecture Search
This work addresses the problem of understanding and potentially improving NAS algorithms for researchers in automated machine learning, though it is incremental as it builds on existing methods.
The paper reveals that existing neural architecture search (NAS) algorithms like DARTS and ENAS favor wide, shallow cell structures due to fast convergence from smooth loss landscapes and accurate gradients, but these may not generalize best, suggesting potential for algorithm improvement.
Neural architecture search (NAS) searches architectures automatically for given tasks, e.g., image classification and language modeling. Improving the search efficiency and effectiveness have attracted increasing attention in recent years. However, few efforts have been devoted to understanding the generated architectures. In this paper, we first reveal that existing NAS algorithms (e.g., DARTS, ENAS) tend to favor architectures with wide and shallow cell structures. These favorable architectures consistently achieve fast convergence and are consequently selected by NAS algorithms. Our empirical and theoretical study further confirms that their fast convergence derives from their smooth loss landscape and accurate gradient information. Nonetheless, these architectures may not necessarily lead to better generalization performance compared with other candidate architectures in the same search space, and therefore further improvement is possible by revising existing NAS algorithms.