Answers are grounded in this paper and cite their sources.
Performance prediction is a critical issue in neural architecture search (NAS), enabling efficient evaluation of candidate architectures without full training. Conventional predictors typically require large numbers of labeled architectures from the target domain, resulting in high computational costs. Knowledge transfer has thus been explored to alleviate this burden, but existing methods primarily focus on either cross-task transfer or cross-space transfer, while neglecting scenarios involving simultaneous variation in both. To address this limitation, in this article, we propose an Unsupervised Adaptive Transferable Predictor (UATP), a label-free approach applicable across diverse tasks and search spaces. UATP integrates a dual-branch encoder with a contrastive objective (CO) to jointly capture the topological and semantic characteristics of architectures, incorporates a domain-adversarial module to mitigate cross-domain discrepancies, and adopts a ranking-aware predictor to achieve accurate performance estimation in unlabeled target domains. Experiments on well-known spaces, i.e., NAS-Bench-101, NAS-Bench-201, and DARTS, across multiple vision benchmark tasks demonstrate that the proposed UATP consistently achieves superior Kendall's Tau (KTau) correlations over existing state-of-the-art predictors under unsupervised settings, highlighting its strong generalization across heterogeneous tasks and search spaces and establishing a scalable, annotation-free solution for neural architecture evaluation.
No takes yet. Share an insight, caveat, or question.
Wang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: