[3Yin2] Interactive session 1

[3Yin2-04] Analyzing the submodule structure in deep learning models to reduce the search space of Neural Architecture Search

〇Mamoru Tomiyama¹, Yusuke Hiki², Yusuke Shibuya², Yuta Tokuoka², Takahiro G. Yamada¹, Akira Funahashi¹ (1.Department of Biosciences and Informatics, Keio University, 2.Graduate School of Science and Technology, Keio University)

Keywords:Neural Architecture Search, Graph Neural Network, Neural Network, Explainability

Neural Architecture Search (NAS) is a method to automatically design architectures of deep learning. In NAS, the optimal architecture is determined from search space, which is a set of architecture substructures. However, the search space of the existing NAS is limited due to the huge amount of computation, so there is a possibility of missing high-performance architectures. Therefore, we hypothesized architectures with task-specific substructures are high-performance, and proposed a method to identify task-specific substructures that contribute to design efficient search space in NAS. We developed a method based on Graph Convolutional Network outputs task specificity for each substructure. We confirmed the correlation between task specificity and task accuracy for each substructure, so our method can identify task-specific substructures that contribute to accuracy. Our method is expected to contribute to NAS development that determines high-performance architectures with reduced computational complexity.