AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

Last Layer Re-Training is Sufficient for Robustness to Spurious Correlations

Polina Kirichenko, Pavel Izmailov, Andrew Gordon Wilson

ICLR 2023 · Architectures and efficiency › RNNs and normalization · approach: Representation geometry and collapse · 32 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
Retrain the last layer on group-balanced validation dataERM feature extractor · Waterbirdsworst-group accuracy: 74.9 -> 92.9Improved
Retrain the last layer on group-balanced validation dataERM feature extractor · CelebAworst-group accuracy: 46.9 -> 88.3Improved
Retrain the last layer on group-balanced validation dataERM feature extractor · MultiNLIworst-group accuracy: 65.9 -> 74.7Improved
Retrain the last layer on group-balanced validation dataERM feature extractor · CivilCommentsworst-group accuracy: 55.6 -> 70.1Improved
Use held-out validation data rather than training data for last-layer retraining (ablation)DFR · Waterbirdsworst-group accuracy: 90.2 -> 92.9Improved
Use held-out validation data rather than training data for last-layer retraining (ablation)DFR · CelebAworst-group accuracy: 80.7 -> 88.3Improved
Retrain only the last linear layer instead of the last two layers (ablation)DFR · Waterbirdsworst-group accuracy: 93.1 -> 93.1No clear effect
Retrain only the last linear layer instead of the last residual block (ablation)DFR · Waterbirdsworst-group accuracy: 90.7 -> 93.1Improved
Retrain only the last linear layer instead of all layers (ablation)DFR · Waterbirdsworst-group accuracy: 35.6 -> 93.1Improved
Retrain only the last linear layer instead of the last two layers (ablation)DFR · CelebAworst-group accuracy: 87.2 -> 88.3Improved
Retrain only the last linear layer instead of the last residual block (ablation)DFR · CelebAworst-group accuracy: 83.0 -> 88.3Improved
Retrain only the last linear layer instead of all layers (ablation)DFR · CelebAworst-group accuracy: 77.8 -> 88.3Improved

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