AI Paper Atlas

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

Causally motivated multi-shortcut identification and removal

Jiayun Zheng, Maggie Makar

NeurIPS 2022 · Label-efficient and robust learning › Recognition and long-tail · approach: Group-robust classification · 0 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
reduce X to low-dimensional sufficient statistic s(X) before conditional independence tests (ablation)Shortcut identification · Waterbirdssimulations identifying both true shortcuts: 1 out of the 10 -> all 10Improved
identify sufficient shortcut subset Vp then apply HSIC penalty and weights (ours) instead of L2 onlyResNet-50 predictor · Waterbirds flipped distributionAUROCImproved
replace L2 penalty with HSIC penalty in weighted models (ablation)W-L2 variants · WaterbirdsAUROCImproved
run HSIC on full auxiliary set Vd or raw X instead of identified Vp (ablation)W-HSIC · WaterbirdsAUROCHurt
apply our method instead of L2 baselineInception-V3 · Diabetic retinopathy flippedAUROC: 0.69 -> 0.72Improved
apply our method instead of W-L2Inception-V3 · Diabetic retinopathy flippedAUROC: 0.68 -> 0.72Improved
apply our method instead of L2 baselineInception-V3 · Diabetic retinopathy ideal distributionAUROC: 0.82 -> 0.83No clear effect
apply our method instead of L2 baselineInception-V3 · Diabetic retinopathy same as trainingAUROC: 0.92 -> 0.91Hurt

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