mlsanityData to 7 Oct 2026

Distilling Model Failures as Directions in Latent Space

arXiv 2206.14754 ↗First posted 29 Jun 2022cs.LGcited by 82

Not on the map. Why?

Abstract (arXiv)

Existing methods for isolating hard subpopulations and spurious correlations in datasets often require human intervention. This can make these methods labor-intensive and dataset-specific. To address these shortcomings, we present a scalable method for automatically distilling a model's failure modes. Specifically, we harness linear classifiers to identify consistent error patterns, and, in turn, induce a natural representation of these failure modes as directions within the feature space. We demonstrate that this framework allows us to discover and automatically caption challenging subpopulations within the training dataset. Moreover, by combining our framework with off-the-shelf diffusion models, we can generate images that are especially challenging for the analyzed model, and thus can be used to perform synthetic data augmentation that helps remedy the model's failure modes. Code available at https://github.com/MadryLab/failure-directions

Key facts. Cited by 82 papers in the map (data through 7 Oct 2026). Builds on 68 of its 100 references that are in the map. First posted on arXiv 29 Jun 2022.

Builds on

68 of 100 references.

References · 68

  1. 2021Learning Transferable Visual Models From Natural Language Supervisioncited by 33,456
  2. 2021High-Resolution Image Synthesis with Latent Diffusion Modelscited by 18,694
  3. 2014Explaining and Harnessing Adversarial Examplescited by 7,287
  4. 2022Hierarchical Text-Conditional Image Generation with CLIP Latentscited by 6,539
  5. 2017Towards Deep Learning Models Resistant to Adversarial Attackscited by 6,326
  6. 2022Photorealistic Text-to-Image Diffusion Models with Deep Language Understandingcited by 6,322
  7. 2015Rethinking the Inception Architecture for Computer Visioncited by 5,451
  8. 2013Intriguing properties of neural networkscited by 5,114
  9. 2016"Why Should I Trust You?": Explaining the Predictions of Any Classifiercited by 5,040
  10. 2014Deep Learning Face Attributes in the Wildcited by 3,679
  11. 2017Axiomatic Attribution for Deep Networkscited by 3,350
  12. 2013Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Mapscited by 2,348
38 more
  1. 2016Equality of Opportunity in Supervised Learningcited by 2,000
  2. 2020Shortcut Learning in Deep Neural Networkscited by 1,928
  3. 2019A Survey on Bias and Fairness in Machine Learningcited by 1,911
  4. 2020The Many Faces of Robustness: A Critical Analysis of Out-of-Distribution Generalizationcited by 1,842
  5. 2019Invariant Risk Minimizationcited by 1,832
  6. 2017Understanding Black-box Predictions via Influence Functionscited by 1,714
  7. 2018ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustnesscited by 1,430
  8. 2018Datasheets for Datasetscited by 1,371
  9. 2020WILDS: A Benchmark of in-the-Wild Distribution Shiftscited by 1,366
  10. 2017Interpretability Beyond Feature Attribution: Quantitative Testing with Concept Activation Vectors (TCAV)cited by 1,309
  11. 2019Distributionally Robust Neural Networks for Group Shifts: On the Importance of Regularization for Worst-Case Generalizationcited by 1,232
  12. 2018Sanity Checks for Saliency Mapscited by 1,149
  13. 2017Counterfactual Fairnesscited by 802
  14. 2014Certifying and removing disparate impactcited by 715
  15. 2017Network Dissection: Quantifying Interpretability of Deep Visual Representationscited by 703
  16. 2019Data Shapley: Equitable Valuation of Data for Machine Learningcited by 617
  17. 2017Interpretable Explanations of Black Boxes by Meaningful Perturbationcited by 611
  18. 2021Just Train Twice: Improving Group Robustness without Training Group Informationcited by 522
  19. 2017CheXNet: Radiologist-Level Pneumonia Detection on Chest X-Rays with Deep Learningcited by 459
  20. 2019Does Learning Require Memorization? A Short Tale about a Long Tailcited by 452
  21. 2021Pervasive Label Errors in Test Sets Destabilize Machine Learning Benchmarkscited by 388
  22. 2016Fairness Beyond Disparate Treatment & Disparate Impact: Learning Classification without Disparate Mistreatmentcited by 368
  23. 2018Fairness Without Demographics in Repeated Loss Minimizationcited by 363
  24. 2020Noise or Signal: The Role of Image Backgrounds in Object Recognitioncited by 342
  25. 2020Are we done with ImageNet?cited by 329
  26. 2019Counterfactual Visual Explanationscited by 312
  27. 2019Don't Take the Easy Way Out: Ensemble Based Methods for Avoiding Known Dataset Biasescited by 279
  28. 2018Explanations based on the Missing: Towards Contrastive Explanations with Pertinent Negativescited by 261
  29. 2017Real Time Image Saliency for Black Box Classifierscited by 246
  30. 2018Failing Loudly: An Empirical Study of Methods for Detecting Dataset Shiftcited by 240
  31. 2019On Completeness-aware Concept-Based Explanations in Deep Neural Networkscited by 239
  32. 2018Multiaccuracy: Black-Box Post-Processing for Fairness in Classificationcited by 221
  33. 2020No Subclass Left Behind: Fine-Grained Robustness in Coarse-Grained Classification Problemscited by 208
  34. 2022Datamodels: Predicting Predictions from Training Datacited by 186
  35. 2017Visualizing Deep Neural Network Decisions: Prediction Difference Analysiscited by 166
  36. 2019Towards Fairer Datasets: Filtering and Balancing the Distribution of the People Subtree in the ImageNet Hierarchycited by 156
  37. 2020BREEDS: Benchmarks for Subpopulation Shiftcited by 155
  38. 2022Domino: Discovering Systematic Errors with Cross-Modal Embeddingscited by 155

50 most cited of 68 All 68 (JSON)

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2022: 8; 2023: 30; 2024: 20; 2025: 18; 2026: 6.

82 papers.

Sort:

By area:

  1. 2022Toward Transparent AI: A Survey on Interpreting the Inner Structures of Deep Neural Networkscited by 126
  2. 2023Diversify Your Vision Datasets with Automatic Diffusion-Based Augmentationcited by 100
  3. 2022A Whac-A-Mole Dilemma: Shortcuts Come in Multiples Where Mitigating One Amplifies Otherscited by 85
  4. 2023Dream the Impossible: Outlier Imagination with Diffusion Modelscited by 77
  5. 2023Discover and Cure: Concept-aware Mitigation of Spurious Correlationcited by 71
  6. 2023Text-To-Concept (and Back) via Cross-Model Alignmentcited by 63
  7. 2023Discovering and Mitigating Visual Biases through Keyword Explanationcited by 58
  8. 2023Diagnosing and Rectifying Vision Models using Languagecited by 58
  9. 2025Interpreting CLIP with Hierarchical Sparse Autoencoderscited by 54
  10. 2023Describing Differences in Image Sets with Natural Languagecited by 51
  11. 2023Mitigating Spurious Correlations in Multi-modal Models during Fine-tuningcited by 47
  12. 2023LANCE: Stress-testing Visual Models by Generating Language-guided Counterfactual Imagescited by 46
38 more
  1. 2022Discovering Bugs in Vision Models using Off-the-shelf Image Generation and Captioningcited by 45
  2. 2022Adaptive Testing of Computer Vision Modelscited by 39
  3. 2023Mass-Producing Failures of Multimodal Systems with Language Modelscited by 36
  4. 2023Dataset Interfaces: Diagnosing Model Failures Using Controllable Counterfactual Generationcited by 35
  5. 2023PUG: Photorealistic and Semantically Controllable Synthetic Data for Representation Learningcited by 31
  6. 2022ModelDiff: A Framework for Comparing Learning Algorithmscited by 30
  7. 2024Multilinear Mixture of Experts: Scalable Expert Specialization through Factorizationcited by 28
  8. 2023FACTS: First Amplify Correlations and Then Slice to Discover Biascited by 26
  9. 2023Red Teaming Deep Neural Networks with Feature Synthesis Toolscited by 21
  10. 2023Effective Human-AI Teams via Learned Natural Language Rules and Onboardingcited by 20
  11. 2022"Why did the Model Fail?": Attributing Model Performance Changes to Distribution Shiftscited by 18
  12. 2023Identification of Systematic Errors of Image Classifiers on Rare Subgroupscited by 17
  13. 2024Slicing Through Bias: Explaining Performance Gaps in Medical Image Analysis using Slice Discovery Methodscited by 14
  14. 2023LOVM: Language-Only Vision Model Selectioncited by 12
  15. 2024Discover and Mitigate Multiple Biased Subgroups in Image Classifierscited by 11
  16. 2024Simplicity Bias of Two-Layer Networks beyond Linearly Separable Datacited by 11
  17. 2024Failures Are Fated, But Can Be Faded: Characterizing and Mitigating Unwanted Behaviors in Large-Scale Vision and Language Modelscited by 11
  18. 2025HiBug2: Efficient and Interpretable Error Slice Discovery for Comprehensive Model Debuggingcited by 11
  19. 2023PRIME: Prioritizing Interpretability in Failure Mode Extractioncited by 9
  20. 2023Concept Distillation: Leveraging Human-Centered Explanations for Model Improvementcited by 9
  21. 2024Revisiting Confidence Estimation: Towards Reliable Failure Predictioncited by 9
  22. 2023Model Selection of Anomaly Detectors in the Absence of Labeled Validation Datacited by 8
  23. 2023GeNIe: Generative Hard Negative Images Through Diffusioncited by 8
  24. 2024Clarify: Improving Model Robustness With Natural Language Correctionscited by 8
  25. 2024Distributionally Generative Augmentation for Fair Facial Attribute Classificationcited by 8
  26. 2024Controlled Training Data Generation with Diffusion Modelscited by 8
  27. 2024DISCERN: Decoding Systematic Errors in Natural Language for Text Classifierscited by 8
  28. 2023Error Discovery by Clustering Influence Embeddingscited by 7
  29. 2025When does a predictor know its own loss?cited by 7
  30. 2025Severing Spurious Correlations with Data Pruningcited by 7
  31. 2022Diagnostics for Deep Neural Networks with Automated Copy/Paste Attackscited by 6
  32. 2023Interpreting CLIP: Insights on the Robustness to ImageNet Distribution Shiftscited by 6
  33. 2024DECIDER: Leveraging Foundation Model Priors for Improved Model Failure Detection and Explanationcited by 6
  34. 2022Identifying the Context Shift between Test Benchmarks and Production Datacited by 5
  35. 2024A hierarchical decomposition for explaining ML performance discrepanciescited by 5
  36. 2024Data Debiasing with Datamodels (D3M): Improving Subgroup Robustness via Data Selectioncited by 5
  37. 2024What could go wrong? Discovering and describing failure modes in computer visioncited by 5
  38. 2025VISLIX: An XAI Framework for Validating Vision Models with Slice Discovery and Analysiscited by 5

50 most cited of 82 All 82 (JSON)

Similar, not cited

Close in wording, no citation either way

  1. 2025FailureAtlas:Mapping the Failure Landscape of T2I Models via Active Exploration
  2. 2025How does My Model Fail? Automatic Identification and Interpretation of Physical Plausibility Failure Modes with Matryoshka Transcoders
  3. 2023Discovering Failure Modes of Text-guided Diffusion Models via Adversarial Search
  4. 2026Quantifying Error Propagation and Model Collapse in Diffusion Models
  5. 2026Tracing Agentic Failure from the Flow of Success
  6. 2026Feature Selective Model Collapse in Diffusion Models: Total Replacement versus Fixed-Budget Training
  7. 2024Difficulty Controlled Diffusion Model for Synthesizing Effective Training Data
  8. 2025MGD3: Mode-Guided Dataset Distillation using Diffusion Models