mlsanityData to 7 Oct 2026

Better Hit the Nail on the Head than Beat around the Bush: Removing Protected Attributes with a Single Projection

arXiv 2212.04273 ↗First posted 8 Dec 2022cs.LG cs.CYcited by 17

Not on the map. Why?

Abstract (arXiv)

Bias elimination and recent probing studies attempt to remove specific information from embedding spaces. Here it is important to remove as much of the target information as possible, while preserving any other information present. INLP is a popular recent method which removes specific information through iterative nullspace projections. Multiple iterations, however, increase the risk that information other than the target is negatively affected. We introduce two methods that find a single targeted projection: Mean Projection (MP, more efficient) and Tukey Median Projection (TMP, with theoretical guarantees). Our comparison between MP and INLP shows that (1) one MP projection removes linear separability based on the target and (2) MP has less impact on the overall space. Further analysis shows that applying random projections after MP leads to the same overall effects on the embedding space as the multiple projections of INLP. Applying one targeted (MP) projection hence is methodologically cleaner than applying multiple (INLP) projections that introduce random effects.

Key facts. Cited by 17 papers in the map (data through 7 Oct 2026). Builds on 27 of its 39 references that are in the map. First posted on arXiv 8 Dec 2022.

Builds on

27 of 39 references.

References · 27

  1. 2018BERT: Pre-training of Deep Bidirectional Transformers for Language Understandingcited by 40,204
  2. 2016Equality of Opportunity in Supervised Learningcited by 2,000
  3. 2016Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddingscited by 1,304
  4. 2016Semantics derived automatically from language corpora contain human-like biasescited by 894
  5. 2016Bag of Tricks for Efficient Text Classificationcited by 869
  6. 2018Mitigating Unwanted Biases with Adversarial Learningcited by 663
  7. 2018Gender Bias in Coreference Resolution: Evaluation and Debiasing Methodscited by 657
  8. 2020Null It Out: Guarding Protected Attributes by Iterative Nullspace Projectioncited by 442
  9. 2018Learning Adversarially Fair and Transferable Representationscited by 353
  10. 2019Bias in Bios: A Case Study of Semantic Representation Bias in a High-Stakes Settingcited by 346
  11. 2019Lipstick on a Pig: Debiasing Methods Cover up Systematic Gender Biases in Word Embeddings But do not Remove Themcited by 268
  12. 2014SimLex-999: Evaluating Semantic Models with (Genuine) Similarity Estimationcited by 170
15 more
  1. 2018Learning Gender-Neutral Word Embeddingscited by 169
  2. 2021Harms of Gender Exclusivity and Challenges in Non-Binary Representation in Language Technologiescited by 128
  3. 2022Linear Adversarial Concept Erasurecited by 98
  4. 2019Understanding Undesirable Word Embedding Associationscited by 59
  5. 2019What's in a Name? Reducing Bias in Bios without Access to Protected Attributescited by 51
  6. 2020OSCaR: Orthogonal Subspace Correction and Rectification of Biases in Word Embeddingscited by 40
  7. 2022Gold Doesn't Always Glitter: Spectral Removal of Linear and Nonlinear Guarded Attribute Informationcited by 33
  8. 2022Can Transformer be Too Compositional? Analysing Idiom Processing in Neural Machine Translationcited by 28
  9. 2020Exploring the Linear Subspace Hypothesis in Gender Bias Mitigationcited by 25
  10. 2022Unit Testing for Concepts in Neural Networkscited by 23
  11. 2022How Conservative are Language Models? Adapting to the Introduction of Gender-Neutral Pronounscited by 16
  12. 2022Analyzing Gender Representation in Multilingual Modelscited by 10
  13. 2021The Rediscovery Hypothesis: Language Models Need to Meet Linguisticscited by 7
  14. 2018Firearms and Tigers are Dangerous, Kitchen Knives and Zebras are Not: Testing whether Word Embeddings Can Tellcited by 5
  15. 2021Obstructing Classification via Projectioncited by 5

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2023: 1; 2024: 5; 2025: 7; 2026: 4.

17 papers.

Sort:

By area:

  1. 2024Refusal in Language Models Is Mediated by a Single Directioncited by 960
  2. 2023LEACE: Perfect linear concept erasure in closed formcited by 254
  3. 2025Convergent Linear Representations of Emergent Misalignmentcited by 66
  4. 2024On Behalf of the Stakeholders: Trends in NLP Model Interpretability in the Era of LLMscited by 24
  5. 2025COSMIC: Generalized Refusal Direction Identification in LLM Activationscited by 15
  6. 2025RepIt: Steering Language Models with Concept-Specific Refusal Vectorscited by 10
  7. 2025Fundamental Limits of Perfect Concept Erasurecited by 5
  8. 2024Can Large Language Models (or Humans) Disentangle Text?cited by 3
  9. 2025Preserving Task-Relevant Information Under Linear Concept Removalcited by 3
  10. 2025Quiet Feature Learning in Algorithmic Taskscited by 2
  11. 2025Improving Causal Interventions in Amnesic Probing with Mean Projection or LEACEcited by 2
  12. 2026Suppression Is Not Forgetting: Residual Recoverability in Visual Concept Unlearning for VLMscited by 1
5 more
  1. 2026SkillEval: Decomposing Agent Skill Quality into Interpretable Signalscited by 1
  2. 2024What Happens to a Dataset Transformed by a Projection-based Concept Removal Method?cited by 0
  3. 2024Balancing the Scales: Reinforcement Learning for Fair Classificationcited by 0
  4. 2026Calibrating Post-Training Feature Shifts for LLM Data Contamination Detectioncited by 0
  5. 2026Iterative Erasure Count Is Not an Affine-Invariant Concept Dimensioncited by 0