Interpretability Analysis Method Based on Path Decomposition for Cell-based Neural Architecture Search
Zitu Liu, Kai Zhu, Yue Liu, Qun Liu, Guoyin Wang, Yike Guo · 2021
Neural Architecture Search (NAS) has become an emerging hot research problem in the field of artificial intelligence recently. The cell-based neural architecture search method has flexible search space conversion and good performance, which has become one of the mainstream search methods of NAS. But current cell-based neural architecture search strategies offer little insight into its interpretability (what kind of architectures the NAS algorithms are learning and why they learn these specific architectures). We proposed an Interpretability Analysis Method Based on Path Decomposition (IAM-PD) that combines the cell similarity representation based on path decomposition and neural network architecture evaluation perspectives. More importantly, our method affords interpretability by discovering useful architecture features and their corresponding impact on the architecture performance. Indeed, we demonstrate empirically that our model is capable of identifying useful motifs which can guide the generation of new architectures. Finally, the experiment results on the two search spaces of NAS-Bench-101 and NAS-Bench-201 show that the similarity metric obtained by IAM-PD has better interpretation than the existing methods.