Learning to Rank Ace Neural Architectures via Normalized Discounted Cumulative Gain

TL;DR

This paper introduces AceNAS, a novel NAS ranking method that optimizes NDCG directly, leading to more effective architecture ranking and significant improvements in search efficiency and accuracy.

Contribution

It proposes a new ranking metric-based NAS algorithm, AceNAS, that outperforms existing methods by focusing on top architecture identification and reducing search costs.

Findings

Outperforms state-of-the-art NAS methods in accuracy.

Achieves up to 8x reduction in search cost.

Demonstrates effectiveness across 12 NAS benchmarks.

Abstract

One of the key challenges in Neural Architecture Search (NAS) is to efficiently rank the performances of architectures. The mainstream assessment of performance rankers uses ranking correlations (e.g., Kendall's tau), which pay equal attention to the whole space. However, the optimization goal of NAS is identifying top architectures while paying less attention on other architectures in the search space. In this paper, we show both empirically and theoretically that Normalized Discounted Cumulative Gain (NDCG) is a better metric for rankers. Subsequently, we propose a new algorithm, AceNAS, which directly optimizes NDCG with LambdaRank. It also leverages weak labels produced by weight-sharing NAS to pre-train the ranker, so as to further reduce search cost. Extensive experiments on 12 NAS benchmarks and a large-scale search space demonstrate that our approach consistently outperforms SOTA NAS methods, with up to 3.67% accuracy improvement and 8x reduction on search cost.