Neural architecture search (NAS) typically requires evaluating many candidate models, which is expensive. Zero-cost proxies (ZCPs) offer a faster alternative by estimating performance without training, but individual proxies tend to capture only limited aspects of a model's quality. A new arXiv paper introduces FENAS, a few-shot meta-ensembling framework designed to combine multiple ZCPs so their strengths complement one another.
Rather than relying on a single proxy, FENAS learns to synergize several proxies using a few-shot meta-learning approach. This allows the framework to adapt quickly to new search spaces or tasks with only a small number of actual evaluations, reducing the overall cost of NAS. The authors position FENAS as a way to improve performance estimation accuracy while staying within strict evaluation budgets.
The paper is a single source, so there are no conflicting findings to compare. Its contribution is primarily methodological: a framework that treats zero-cost proxies as an ensemble and uses meta-learning to make the combination efficient. The abstract does not include experimental numbers, so the practical gains remain to be detailed in the full text.