Covariance-Inverse Fusion: Toward Comprehensive Proxies for Zero-Shot Neural Architecture Search

Zero-cost proxy fusion for neural architecture search (NAS) shows potential for efficient architecture evaluation. However, current proxy fusion methods still require extensive ground-truth samples for proxy generation and validation, incurring high computational cost. One of the keys to overcoming this challenge lies in developing automated fusion mechanisms that efficiently integrate the complementary strengths of diverse zero-cost proxies. To address this challenge, this work encodes such complementary information in the inverse covariance matrix. Based on the inverse matrix, the covariance-inverse fusion (CoIF) method is proposed to leverage the complementary strengths of existing proxies through covariance-weighted fusion. Theoretically, the proposed method is proven to provide optimal fusion weights and guarantee nondecreasing consistency. In addition, the variance of the estimated correlation is inversely proportional to the sample size, enabling reliable fusion with low ground-truth dependency. CoIF achieves competitive performance with 80% fewer samples and 99% less fusion time compared to existing learning-based fusion methods. Experiments across convolutional neural networks (CNNs), vision transformers, and large language models (LLMs) demonstrate that CoIF effectively integrates diverse architectural properties through proxy fusion. The closed-form fusion weights ensure both efficiency and theoretical optimality in the linear fusion case, establishing CoIF as an effective approach for zero-shot (ZS) NAS.