Research map / Label-efficient and robust learning / Recognition and long-tail
Removing Spurious Concepts from Neural Network Representations via Joint Subspace Estimation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Joint Subspace Estimation (JSE), an iterative post-hoc concept-removal method that jointly estimates orthogonal spurious and main-task subspaces in DNN embeddings so only spurious directions are removed.
Problem
Post-hoc concept-removal methods remove spurious concepts from embeddings but also inadvertently remove main-task features when the spurious and main-task concepts are correlated, hurting performance and interpretability.
Approach
Jointly estimate orthogonal spurious and main-task concept vectors via a constrained logistic regression, iterate with nested loops that project out estimated directions, and use BCE-based t-tests to decide whether a direction is spurious or main-task, so only the spurious subspace is removed.
Key results
- On Toy data with strong spurious correlation (rho=0.9), JSE keeps worst-group accuracy near 80.3% while ERM drops to 68.4% and INLP to 47.1%.
- On Waterbirds at p_train=0.9, JSE achieves 87.8% worst-group accuracy versus 72.4% for ERM and 71.8% for INLP (per-group OOD evaluation).
- On MultiNLI with a clean BERT model, only JSE avoids performance loss relative to ERM, while INLP, LEACE and RLACE degrade as the spurious correlation increases.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use joint subspace estimation for concept removal instead of existing concept-removal methods | ResNet50 on Waterbirds · Waterbirds OOD test, train correlation 0.9 | worst-group accuracy: 72.40 (0.62) -> 87.77 (0.36) | Helped |
| Use joint subspace estimation for concept removal instead of existing concept-removal methods | ResNet50 on CelebA · CelebA OOD test, train correlation 0.9 | worst-group accuracy: 57.04 (2.06) -> 74.36 (1.51) | Helped |
| Use joint subspace estimation for concept removal instead of the strongest baseline | BERT on MultiNLI · MultiNLI OOD test, train correlation 0.9 | worst-group accuracy: 67.63 (1.11) -> 63.42 (1.26) | Hurt |
| Use joint subspace estimation for concept removal instead of existing concept-removal methods | Toy dataset · Toy OOD test, ρ=0.9 | worst-group accuracy: 68.41 (0.91) -> 80.27 (0.42) | Helped |
| Swap the inner and outer loops used to identify main-task and spurious directions (ablation) | JSE · Experiments reported in Appendix | outcome | No clear effect |
| Train the linear classifier on the main-task subspace instead of the orthogonal complement of the spurious subspace (ablation) | JSE · Experiments reported in Appendix | main-task classification performance | No clear effect |
Surprising findings
- On MultiNLI, JSE performs similarly to the best other concept-removal methods (LEACE and RLACE) rather than clearly better, which the authors attribute to overlapping concepts in BERT [CLS] embeddings.
- On the Toy dataset, JSE's performance still declines at strong spurious correlation, which the authors attribute to finite-sample estimation noise.
Limitations
- Assumes linear subspaces and orthogonality of spurious and main-task subspaces.
- Performance loss under strong spurious correlation due to finite-sample estimation noise.
- Gains are limited on BERT [CLS] embeddings where concepts overlap.
Experimental setup
Datasets: Waterbirds CelebA MultiNLI Toy (synthetic Gaussian embeddings)
Benchmarks: Waterbirds CelebA MultiNLI Toy dataset
Metrics: worst-group accuracy average accuracy mean-squared error of OLS reconstruction of main-task and sp
Baselines: INLP RLACE LEACE ADV ERM group-weighted ERM GDRO JTT
Builds on: iterative null-space projection relaxed linear adversarial concept erasure least-squares concept erasure Grad-CAM
Closest papers by meaning
- Let Samples Speak: Mitigating Spurious Correlation by Exploiting the Clusterness of Samples (CVPR 2025 · similarity 0.84)
- Severing Spurious Correlations with Data Pruning (ICLR 2025 · similarity 0.84)
- Preserving Task-Relevant Information Under Linear Concept Removal (NeurIPS 2025 · similarity 0.83)
- Explainable Neural Networks with Guarantee: A Sparse Estimation Approach (AAAI 2025 · similarity 0.83)
- Label-Efficient Group Robustness via Out-of-Distribution Concept Curation (CVPR 2024 · similarity 0.83)
- SASA: Subspace-Aware Sparse Autoencoders for Effective Mechanistic Interpretability (NeurIPS 2026 · similarity 0.82)
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.