Prediction-powered Neural Architecture Search
PPNAS combines scarce architecture labels with zero-cost proxies for label-efficient neural architecture search.
PPNAS applies prediction-powered inference to neural architecture search, combining a small set of architectures with observed performance labels and a much larger set scored by noisy zero-cost proxies. Zero-cost proxies supply ordinal pairwise ranking supervision, and PPI corrects systematic gaps between proxy rankings and true performance. In end-to-end predictor-based NAS, the method reports state-of-the-art results when evaluation budgets are limited. The authors describe it as the first prediction-powered approach to label-efficient NAS.
- PPNAS fuses scarce true labels with large-scale zero-cost proxy scores.
- Ordinal proxy rankings supply pairwise supervision that PPI then debiases.
- End-to-end predictor-based NAS is state-of-the-art under limited budgets.
- Authors call it the first prediction-powered method for label-efficient NAS.
From the original article
1 min read at arxiv.orgEvaluating candidate architectures in neural architecture search (NAS) faces an inherent trade-off: on the one hand, reliable performance labels are limited because training and evaluating architectures is costly; on the other hand, zero-cost proxies (ZCPs) are cheap to compute at large scale but can be noisy. Yet, how to effectively combine these two sources of supervision remains unclear.
By Pascal Janetzky · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.