AI Paper Atlas

Research map / Label-efficient and robust learning / OOD and anomaly detection

Agreement-on-the-line: Predicting the Performance of Neural Networks under Distribution Shift

Christina Baek, Yiding Jiang, Aditi Raghunathan, J Zico Kolter

NeurIPS 2022 · Label-efficient and robust learning › OOD and anomaly detection · approach: Provable and feature-based OOD detection · 10 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
Estimate OOD accuracy by solving for model accuracies using OOD pairwise agreement (ALine-D)Neural-network classifiers · CIFAR-10.1, CIFAR-10.2, ImageNetV2, CIFAR-10C-Fog, CIFAR-10C-Snow, CIFAR-10C-Saturate, fMoW-wilds, RxRx1-wildsMAE (%): 1.11; 3.93; 2.06; 1.45; 1.32; 0.41; 1.30; 0.27Improved
Estimate OOD accuracy by solving for model accuracies using pairwise agreement (ALine-D)Neural-network classifiers · Camelyon17-wilds and iWildCam-wildsMAE (%): 5.47; 4.95Mixed
Use agreement-line slope and bias to transform ID accuracy without directly using the target model's OOD agreement (ALine-S)Neural-network classifiers · CIFAR-10.1, CIFAR-10.2, ImageNetV2, CIFAR-10C-Fog, CIFAR-10C-Snow, CIFAR-10C-Saturate, fMoW-wilds, RxRx1-wildsMAE (%): 1.17; 3.93; 2.08; 1.75; 1.97; 0.77; 1.44; 0.52Mixed
Use ALine-D rather than ALine-SOOD accuracy prediction · CIFAR-10.1, ImageNetV2, CIFAR-10C-Fog, CIFAR-10C-Snow, CIFAR-10C-Saturate, fMoW-wilds, RxRx1-wildsMAE (%): 1.17 -> 1.11; 2.08 -> 2.06; 1.75 -> 1.45; 1.97 -> 1.32; 0.77 -> 0.41; 1.44 -> 1.30; 0.52 -> 0.27Improved
Use ALine-D rather than ATCOOD accuracy prediction · ImageNetV2 and CIFAR-10C-SnowMAE (%): 2.06 vs 1.14; 1.32 vs 1.31Hurt
Use ALine-D rather than ProjNormResNet18 OOD accuracy estimates · CIFAR-10CR²: 0.973 -> 0.974Improved
Use agreement-line estimation on a single training trajectory rather than independently trained modelsResNet18 on CIFAR-10 and ResNet50 on ImageNet · CIFAR-10.1 and ImageNetV2MAE (%): 2.19%; 3.61%Mixed
Use more models in the ALine-D model set (ablation)ALine-D · CIFAR-10.1MAE (%)Improved
Use more models in the ALine-D model set (ablation)ALine-D · ImageNetV2MAE (%): 2.5%Improved
Use more models in the ALine-D model set (ablation)ALine-D · CIFAR-10C FogMAE (%): below 1.8%No clear effect
Use a diverse-architecture model set instead of models with one architecture (ablation)ALine-D · CIFAR-10.1 and CIFAR-10C FogMAE (%)Improved
Use a diverse-architecture model set instead of models with one architecture (ablation)ALine-D · fMoW-wildsMAE (%)Hurt

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