mlsanityData to 7 Oct 2026

Freeze then Train: Towards Provable Representation Learning under Spurious Correlations and Feature Noise

arXiv 2210.11075 ↗First posted 20 Oct 2022cs.LGcited by 22

Not on the map. Why?

Abstract (arXiv)

The existence of spurious correlations such as image backgrounds in the training environment can make empirical risk minimization (ERM) perform badly in the test environment. To address this problem, Kirichenko et al. (2022) empirically found that the core features that are related to the outcome can still be learned well even with the presence of spurious correlations. This opens a promising strategy to first train a feature learner rather than a classifier, and then perform linear probing (last layer retraining) in the test environment. However, a theoretical understanding of when and why this approach works is lacking. In this paper, we find that core features are only learned well when their associated non-realizable noise is smaller than that of spurious features, which is not necessarily true in practice. We provide both theories and experiments to support this finding and to illustrate the importance of non-realizable noise. Moreover, we propose an algorithm called Freeze then Train (FTT), that first freezes certain salient features and then trains the rest of the features using ERM. We theoretically show that FTT preserves features that are more beneficial to test time probing. Across two commonly used spurious correlation datasets, FTT outperforms ERM, IRM, JTT and CVaR-DRO, with substantial improvement in accuracy (by 4.5%) when the feature noise is large. FTT also performs better on general distribution shift benchmarks.

Key facts. Cited by 22 papers in the map (data through 7 Oct 2026). Builds on 52 of its 76 references that are in the map. First posted on arXiv 20 Oct 2022.

Builds on

52 of 76 references.

References · 52

  1. 2015Deep Residual Learning for Image Recognitioncited by 51,809
  2. 2020Language Models are Few-Shot Learnerscited by 32,199
  3. 2021Masked Autoencoders Are Scalable Vision Learnerscited by 7,339
  4. 2014Deep Learning Face Attributes in the Wildcited by 3,679
  5. 2019Invariant Risk Minimizationcited by 1,832
  6. 2019Class-Balanced Loss Based on Effective Number of Samplescited by 1,239
  7. 2019Distributionally Robust Neural Networks for Group Shifts: On the Importance of Regularization for Worst-Case Generalizationcited by 1,232
  8. 2017Deep Hashing Network for Unsupervised Domain Adaptationcited by 1,181
  9. 2020In Search of Lost Domain Generalizationcited by 1,042
  10. 2017Deeper, Broader and Artier Domain Generalizationcited by 983
  11. 2020Out-of-Distribution Generalization via Risk Extrapolation (REx)cited by 737
  12. 2022Fine-Tuning can Distort Pretrained Features and Underperform Out-of-Distributioncited by 722
38 more
  1. 2021Just Train Twice: Improving Group Robustness without Training Group Informationcited by 522
  2. 2017A systematic study of the class imbalance problem in convolutional neural networkscited by 510
  3. 2022Last Layer Re-Training is Sufficient for Robustness to Spurious Correlationscited by 379
  4. 2018Fairness Without Demographics in Repeated Loss Minimizationcited by 363
  5. 2018Learning Adversarially Fair and Transferable Representationscited by 353
  6. 2020The Pitfalls of Simplicity Bias in Neural Networkscited by 349
  7. 2017Implicit Regularization in Matrix Factorizationcited by 318
  8. 2020An Investigation of Why Overparameterization Exacerbates Spurious Correlationscited by 310
  9. 2018Joint Optimization Framework for Learning with Noisy Labelscited by 293
  10. 2018On the Optimization of Deep Networks: Implicit Acceleration by Overparameterizationcited by 280
  11. 2021Understanding self-supervised Learning Dynamics without Contrastive Pairscited by 254
  12. 2018Multiaccuracy: Black-Box Post-Processing for Fairness in Classificationcited by 221
  13. 2020Few-Shot Learning via Learning the Representation, Provablycited by 213
  14. 2022Improving Out-of-Distribution Robustness via Selective Augmentationcited by 202
  15. 2016Demographic Dialectal Variation in Social Media: A Case Study of African-American Englishcited by 194
  16. 2020On the Theory of Transfer Learning: The Importance of Task Diversitycited by 186
  17. 2018Algorithmic Regularization in Learning Deep Homogeneous Models: Layers are Automatically Balancedcited by 172
  18. 2022Correct-N-Contrast: A Contrastive Approach for Improving Robustness to Spurious Correlationscited by 171
  19. 2020Provable Meta-Learning of Linear Representationscited by 161
  20. 2019Linearized two-layers neural networks in high dimensioncited by 155
  21. 2016Variance-based regularization with convex objectivescited by 153
  22. 2020Predicting What You Already Know Helps: Provable Self-Supervised Learningcited by 152
  23. 2019Distributionally Robust Language Modelingcited by 142
  24. 2018Invariance, Causality and Robustnesscited by 123
  25. 2019Implicit Regularization of Discrete Gradient Dynamics in Linear Neural Networkscited by 122
  26. 2017Learning From Noisy Large-Scale Datasets With Minimal Supervisioncited by 115
  27. 2021Learning with Feature-Dependent Label Noise: A Progressive Approachcited by 102
  28. 2022Agree to Disagree: Diversity through Disagreement for Better Transferabilitycited by 69
  29. 2018A Continuous-Time View of Early Stopping for Least Squarescited by 67
  30. 2021Examining and Combating Spurious Features under Distribution Shiftcited by 64
  31. 2022C-Mixup: Improving Generalization in Regressioncited by 57
  32. 2021The Power of Contrast for Feature Learning: A Theoretical Analysiscited by 53
  33. 2021Meta-Learning with Fewer Tasks through Task Interpolationcited by 45
  34. 2021Adversarial Training Helps Transfer Learning via Better Representationscited by 36
  35. 2021Out-of-distribution Prediction with Invariant Risk Minimization: The Limitation and An Effective Fixcited by 27
  36. 2022Nearly Minimax Algorithms for Linear Bandits with Shared Representationcited by 16
  37. 2017Using SVD for Topic Modelingcited by 14
  38. 2021Understanding Dynamics of Nonlinear Representation Learning and Its Applicationcited by 11

50 most cited of 52 All 52 (JSON)

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2023: 7; 2024: 9; 2025: 4; 2026: 2.

22 papers.

Sort:

By area:

  1. 2024TFG: Unified Training-Free Guidance for Diffusion Modelscited by 103
  2. 2024Calibrated Self-Rewarding Vision Language Modelscited by 77
  3. 2023Discover and Cure: Concept-aware Mitigation of Spurious Correlationcited by 71
  4. 2023Understanding Multimodal Contrastive Learning and Incorporating Unpaired Datacited by 43
  5. 2023Robust Learning with Progressive Data Expansion Against Spurious Correlationcited by 39
  6. 2023Understanding and Improving Feature Learning for Out-of-Distribution Generalizationcited by 35
  7. 2023HappyMap: A Generalized Multi-calibration Methodcited by 17
  8. 2024Fine-Grained Verifiers: Preference Modeling as Next-token Prediction in Vision-Language Alignmentcited by 17
  9. 2024Complexity Matters: Dynamics of Feature Learning in the Presence of Spurious Correlationscited by 12
  10. 2024Sharpness-Aware Minimization Enhances Feature Quality via Balanced Learningcited by 9
  11. 2024Constructing Concept-based Models to Mitigate Spurious Correlations with Minimal Human Effortcited by 7
  12. 2025Severing Spurious Correlations with Data Pruningcited by 7
10 more
  1. 2024Towards Formalizing Spuriousness of Biased Datasets Using Partial Information Decompositioncited by 6
  2. 2024BMFT: Achieving Fairness via Bias-based Weight Masking Fine-tuningcited by 6
  3. 2024Diffexplainer: Towards Cross-modal Global Explanations with Diffusion Modelscited by 5
  4. 2025Sparse Model Inversion: Efficient Inversion of Vision Transformers for Data-Free Applicationscited by 4
  5. 2023FedBug: A Bottom-Up Gradual Unfreezing Framework for Federated Learningcited by 3
  6. 2025On the Unreasonable Effectiveness of Last-layer Retrainingcited by 3
  7. 2025NeuronTune: Towards Self-Guided Spurious Bias Mitigationcited by 2
  8. 2023Hierarchical Simplicity Bias of Neural Networkscited by 0
  9. 2026Whitening Improves Robustness to Spurious Correlations in Linear Probescited by 0
  10. 2026The Conflict Between Logic and Memory: Training Conditions for Optimizer-Dependent Rule Acquisitioncited by 0