RiPPLE: Cross-Space Performance Prediction from Early Training for Neural Architecture Search

Abstract

Neural architecture search evaluates candidate networks, and the most reliable way to measure one is to train it fully, which is expensive, so evaluating an entire space this way is prohibitive. Two cheaper families each fall short. Zero-cost proxies score every architecture at initialization, but no single proxy holds across spaces, and even the strongest leaves a wide gap from a faithful ranking. Fully supervised predictors are more accurate, yet every label is a full training of a sampled architecture, and those adding a partial-training signal read it as a per-candidate feature, so cost grows with the space. Our goal is to rank an entire space well at a fractional budget. We introduce RiPPLE, Ranking via Prefix-Propagated Label Extrapolation, which obtains its labels cheaply: it trains a small coverage set of anchors to an early prefix, extrapolates each learning curve to a surrogate label, and propagates these labels over label-free features. The signal stays a label on the anchors, never a per-candidate feature, and no per-space choice reads held-out accuracy, so one unchanged pipeline ranks the spaces at fractional budgets. Across twelve cells from four search-space families, RiPPLE ranks near the oracle, matching or surpassing predictors under a common budget ceiling and holding where zero-cost proxies collapse; read as a selector, its chosen architectures are competitive with leading search.