AI Paper Atlas

Research map / Label-efficient and robust learning / Recognition and long-tail

Beyond Confidence: Reliable Models Should Also Consider Atypicality

Mert Yuksekgonul, Linjun Zhang, James Zou, Carlos Guestrin

NeurIPS 2023 · Label-efficient and robust learning › Recognition and long-tail · approach: Confidence calibration objectives · 2 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
Add input atypicality-dependent recalibration and class-specific correctionsImageNet ResNet152 · ImageNet, atypicality quantile 1.0 (GMM)ECE: 0.033 ± 0.004 -> 0.026 ± 0.001Improved
Add input atypicality-dependent recalibration and class-specific correctionsImageNet ResNet18 · ImageNet, atypicality quantile 1.0 (GMM)ECE: 0.032 ± 0.003 -> 0.017 ± 0.002Improved
Add input atypicality-dependent recalibration and class-specific correctionsCIFAR10 WideResNet28 · CIFAR10, atypicality quantile 1.0 (GMM)ECE: 0.046 ± 0.002 -> 0.041 ± 0.002Improved
Add input atypicality-dependent recalibration and class-specific correctionsCIFAR100 WideResNet28 · CIFAR100, atypicality quantile 1.0 (GMM)ECE: 0.101 ± 0.004 -> 0.109 ± 0.005Hurt
Use atypicality-aware rather than ordinary APS thresholdsAPS on ResNet50 · ImageNet, fitted to 95% coverageaverage prediction set size: 21.3 -> 15.6Improved
Use atypicality-aware rather than ordinary RAPS thresholdsRAPS on ResNet50 · ImageNet, fitted to 95% coverageaverage prediction set size: 4.2 -> 9.1Hurt
Use atypicality-aware prediction-set thresholdsAPS and RAPS · ImageNet high-atypicality and low-confidence groupsconditional coverageImproved
Recalibrate with atypicality without using group annotationsSkin lesion classifier · Fitzpatrick17k skin-tone groupsworst-group accuracy: 0.69 vs 0.63Improved
Condition temperature scaling on atypicality groups instead of using one temperatureTemperature Scaling · Atypicality groupscalibrationMixed
Use atypicality to stratify prediction-quality analysisDiscriminative classifiers and Alpaca-7B · Balanced, imbalanced, and LLM classificationaccuracyHurt
Use atypicality to stratify prediction-quality analysisDiscriminative classifiers and Alpaca-7B · Balanced, imbalanced, and LLM classificationoverconfidenceHurt

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy