CoRA-NAS: Coarse Ranking and Anchor-Residual Refinement for Neural Architecture Search
While zero-cost proxies evaluate neural architectures with minimal overhead, their ranking reliability frequently degrades across different search spaces. CoRA-NAS counters this instability through a two-stage approach: it establishes a static ranking prior by aggregating initialization proxies, then samples anchor models and extrapolates their early validation curves to propagate residual corrections via an ExtraTrees model. Requiring roughly 1% of full training compute without relying on ground-truth training labels, CoRA-NAS achieves mean Spearman correlations between 0.715 and 0.946 across four standard NAS benchmarks, selecting top architectures within 0.05% of the benchmark ceiling.
Why it's worth reading
CoRA-NAS tackles the cross-space inconsistency of zero-cost proxies by anchoring early learning curves at just 1% compute overhead, offering a robust and practical middle ground for architecture search.