AI Paper Atlas

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

When Does Group Invariant Learning Survive Spurious Correlations?

Yimeng Chen, Ruibin Xiong, Zhi-Ming Ma, Yanyan Lan

NeurIPS 2022 · Label-efficient and robust learning › Recognition and long-tail · approach: Data augmentation and generalization bias · 4 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
replace EIIL groups and objective with SCILL (IRM penalty)group invariant learning · PC-MNIST test, ID selectionaccuracy: 50.63 -> 62.49Improved
replace EIIL with SCILL (REx), TEV selectiongroup invariant learning · PC-MNIST testaccuracy: 51.01 -> 67.33Improved
replace EIIL with SCILL (REx)BERT base · MNLI-HANS test, ID selectionaccuracy: 65.16 -> 69.73Improved
replace EIIL with SCILL (cMMD)BERT base · MNLI-HANS test, ID selectionaccuracy: 63.22 -> 69.15Improved
SCILL-REx with ID selection vs PoE and ConfRegBERT base · MNLI-HANSaccuracy: 69.2 -> 69.73Improved
remove instance reweighting (SCILL uw) (ablation)SCILL-IRM · PC-MNIST test, Oracle selectionaccuracy: 67.46 -> 53.51Hurt
remove instance reweighting (SCILL uw) (ablation)SCILL-IRM · MNLI-HANS test, ID selectionaccuracy: 69.11 -> 64.30Hurt
use majority/minority groups instead of ground-truth strata (ablation)IRM · PC-MNIST test, Oracle selectionaccuracy: 70.19 -> 63.85Hurt
unbalance label weights (p_err 1 -> 2) (ablation)SCILL-gt-IRM · PC-MNIST test, TEV selectionaccuracy: 70.91 -> 56.20Hurt
unbalance label weights (p_err 1 -> 1.2) (ablation)SCILL-gt-IRM · PC-MNIST test, TEV selectionaccuracy: 70.91 -> 63.43Hurt
lower statistical split threshold from 10 to 5 (ablation)SCILL-IRM · PC-MNIST test, ID selectionaccuracy: 62.49 -> 55.50Hurt
raise threshold from 10 to 20 (ablation)SCILL-IRM · PC-MNIST test, Oracle selectionaccuracy: 67.46 -> 67.95No clear effect

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