Inter-layer Transition in Neural Architecture Search

Differential Neural Architecture Search (NAS) methods represent the network architecture as a repetitive proxy directed acyclic graph (DAG) and optimize weights alternatively in differential manner. However, existing model on each edge (i.e., layer network) statistically independent variables, ignoring dependency between edges DAG induced by their topological connections. In this paper, we make first attempt to investigate such proposing novel Inter-layer Transition NAS method. It casts optimization into sequential decision process where of connected is explicitly modeled. Specifically, are divided inner outer groups according whether or not predecessor same cell. While optimized independently, those derived sequentially based learnable transition matrices an attentive probability Experiments five benchmarks confirm value modeling inter-layer demonstrate proposed method outperforms state-of-the-art methods.