Multi-Task Neural Architecture Search Using Architecture Embedding and Transfer Rank

Multi-task neural architecture search (MTNAS) transfers architectural knowledge on neural models between different tasks, with the basic aim of optimizing neural architectures. MTNAS enables effective architectural patterns contained in the source task to be transferred and reused in the target task, which is known as transferability. However, ranking disorder between the source and target task degrades architecture transferability. To address this issue, we devise an architecture transferability quantifying method (ATQ). First, our data-agnostic method maps neural architectures into architecture embedding vectors for subsequent transferability prediction. Second, we apply transfer rank, an instance-based classifier, to alleviate transferability degradation problem. In this work, ATQ is incorporated into an evolutionary multi-task NAS algorithm (KTNAS), to select candidate architectures with better transferability. Extensive experiments are conducted on multiple benchmark datasets, such as NASBench-201, TransNAS-Bench-101 and DARTs. Experimental results show that KTNAS outperforms peer MTNAS algorithms in search efficiency and downstream task performance. Ablation study demonstrates the vital importance of transfer rank on enhancing transfer performance.