acceptodds
Desk rejected at ICLR 2027

Allocating Additional Labels Between Cascade Predictors and Selectors

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

A useful learned selector does not establish that the next labels should train the selector rather than the predictors. Controlled image-cascade experiments separate these questions. In an adapted CIFAR-100 long-tailed condition, class-conditioned routing improves on a strong scalar rule by 2.206 percentage points (pp), but expanding gate support contributes only 0.200pp. Allocating the identical additional labels to the predictors instead improves the source-selected cascade by 0.881pp; two additional paired training repetitions give 1.612pp. Tiny ImageNet preserves the investment advantage, including after SOURCE-selected answer correction. A complete endpoint-support × sampling-prior factorial identifies a 5.901pp gain from distinct examples at the expanded prior, rising to 7.088pp under a separately fixed fourfold training horizon. Prior-only and routing-interaction contrasts remain unresolved. Component controls connect the support effect to an executable alternative: source-selected upgraded cheap-only prediction exceeds the old Tiny cascade by 2.965pp. Corrected Tiny and stronger CIFAR predictors retain fixed-budget investment gains without crossing the corresponding old-pair hindsight ceilings. When corrected budgets are also SOURCE-selected, cheap-only prediction has lower resident request time, but its same-label accuracy advantage remains unresolved. Thus useful routing, the return to further gate labels, and the answers made available by predictor training are distinct quantities. Reported intervals condition on fitted models and exposed images; the comparisons do not establish a universal allocation rule or a lifecycle speedup.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
budgeted prediction, image cascades, supervision allocation, confidence-based selection, predictor training, dynamic feature acquisition

Related papers

Discussion (0)

Sign in to comment.