Efficient Multi-Fidelity NAS with Zero-Cost Proxy-Guided Local Search

Quan Minh Phan, Ngoc Hoang Luong

Keywords: neural architecture search, local search, zero-cost proxies, local optimal network

Abstract:

Using zero-cost (ZC) metrics as proxies for network performance is currently trendy in Neural Architecture Search (NAS) because the low computing cost of these metrics allows search algorithms to thoroughly explore the architecture search space. Nevertheless, recent studies indicate that relying exclusively on ZC proxies appears to be less effective than using traditional training-based metrics, such as validation accuracy, in seeking high-performance networks. Training-based metrics are preferred as the main search objective to guide search algorithms to approach truly good architectures while ZC proxies could be used as low-cost surrogates to accelerate the search process.
ZC proxies with high rank-correlations to network test accuracy are supposed to bring better search results than metrics with lower correlations. In this study, we investigate the effectiveness of ZC proxies in NAS by taking a deeper look into their fitness landscapes rather than focusing only on rank correlations. We construct fitness landscapes of ZC proxy-based local searches by utilizing the Local Optima Network (LON), which is a powerful visualization tool to analyze combinatorial optimization problems. Our findings exhibit that a high correlation does not guarantee finding high-performance architectures, and ZC proxies with low correlations could still be better in certain situations. Our results further consolidate the suggestion of favoring training-based metrics over ZC proxies as the search objective.
Although we could figure out the architectures having the optimal ZC proxy scores, their true performance is often poor.
We then utilize insights from our landscape analysis to propose $\textbf{M}$ulti-$\textbf{F}$idelity $\textbf{N}$eural $\textbf{A}$rchitecture $\textbf{S}$earch (MF-NAS), which is a novel framework that makes use of the efficiency of ZC proxies and the efficacy of training-based metrics. Experimental results on a wide range of NAS benchmarks, i.e., NAS-Bench-101, NAS-Bench-201, and NAS-Bench-ASR, demonstrate the superiority of our proposed approach to state-of-the-art NAS methods under a strict budget.

Primary Area: general machine learning (i.e., none of the above)

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 2448