Capacity Meters That Change Sign: Diagnosing Zero-Cost NAS Proxy Transport Across Search Spaces

Zhenyu Liu, Zheng Gao, Xiaoyu Li, Zhiyuan Jiang, Yang Song, Jiaojiao Jiang

IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), Evaluations and Datasets Track, 2027

Abstract

Zero-cost NAS proxies rank architectures from a single pass at initialization, yet a score that ranks with accuracy in one setting can rank against it in another. Using a 29-feature zoo on the topology and size axes of NAS-Bench-201, we observationally separate a dominant capacity-associated factor from an early-trainability residual. The factor explains 52.9%/49.4% of zoo variation and remains associated with effective rank under raw/global orientation, but the cross-axis loading cosine falls from 0.87 under outcome-informed orientation to 0.46. Several raw proxies also reverse sign, including the condition-number sum (+0.88 → −0.77); dimension and matched-sample controls support a dimension-linked/proxy-geometry-dependent failure, not an independent mechanism, while matched-sample-count resampling preserves the headline reversals. Two controlled search traps incur +5.40 and +1.97 point costs at K=10, while two null directions incur zero. The diagnostic uses cached target-calibration labels and is not fully label-free. These benchmark-bound results motivate validating proxy orientation before transporting a source ranking.

BibTeX

@inproceedings{liu2027capacitymeters,
  title     = {Capacity Meters That Change Sign: Diagnosing Zero-Cost NAS Proxy Transport Across Search Spaces},
  author    = {Liu, Zhenyu and Gao, Zheng and Li, Xiaoyu and Jiang, Zhiyuan and Song, Yang and Jiang, Jiaojiao},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
  note      = {Evaluations and Datasets Track},
  year      = {2027},
}