AI Paper Atlas

Research map / Label-efficient and robust learning / Test-time adaptation

AdaNPC: Exploring Non-Parametric Classifier for Test-Time Adaptation

YiFan Zhang, xue wang, Kexin Jin, Kun Yuan, Zhang Zhang, Liang Wang, Rong Jin, Tieniu Tan

ICML 2023 · Label-efficient and robust learning › Test-time adaptation · approach: Realistic and stable test-time adaptation · 13 citations (OpenAlex)

Read the paper

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
Replace the linear classifier with AdaNPC KNNResNet50 test-time adaptation · PACS, BSZ=32accuracy: 86.68 -> 88.93Improved
Use AdaNPC rather than the strongest listed DG baselinedomain generalization models · OOD benchmarks averageaccuracy: 71.7 -> 72.7Improved
Add BN retraining at test time (ablation)AdaNPC · OOD benchmarks averageaccuracy: 72.7 -> 72.9Improved
Add BN retraining at test time (ablation)AdaNPC · PACS, ResNet50accuracy: 85.8 -> 85.7No clear effect
Train representations with KNN loss rather than ERM cross-entropy (ablation)AdaNPC · PACS test-domain validationaccuracy: 84.2 -> 86.1Improved
Vary the KNN neighbor count and memory size (ablation)KNN-loss AdaNPC · PACS test-domain validationaccuracy: 85.2; 85.1; 86.1; 84.5; 85.7Mixed
Use AdaNPC instead of gradient-updated full-model pseudo-label adaptationResNet50 TTA · PACS, BSZ=2accuracy: 73.35 -> 88.93Improved
Add online target features and predicted labels to memoryAdaNPC · successive adaptation on Rotated MNISTaccuracy: +50%Improved

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