CoRA-NAS: Coarse Ranking and Anchor-Residual Refinement for Neural Architecture Search
CoRA-NAS claims strong cross-space architecture ranking using about 1% of full training cost.
The paper proposes a two-stage neural architecture search framework that first builds a proxy-based ranking, then refines it with early validation curves from sampled anchors. It does not use fully trained architecture-accuracy labels to fit the ranker. Across four NAS benchmarks, CoRA-Refine reports mean Spearman correlations from 0.715 to 0.946, with its worst-space result topping the compared methods. On NAS-Bench-201/CIFAR-100, its selected architecture reaches 73.32% accuracy, just below the reported 73.37% best. ArXiv · AI/CL/LG's note
The paper proposes a two-stage neural architecture search framework that first builds a proxy-based ranking, then refines it with early validation curves from sampled anchors. It does not use fully trained architecture-accuracy labels to fit the ranker. Across four NAS benchmarks, CoRA-Refine reports mean Spearman correlations from 0.715 to 0.946, with its worst-space result topping the compared methods. On NAS-Bench-201/CIFAR-100, its selected architecture reaches 73.32% accuracy, just below the reported 73.37% best. ArXiv · AI/CL/LG's note
score 4