mlsanityData to 7 Oct 2026

Unlearning-based Neural Interpretations

arXiv 2410.08069 ↗First posted 10 Oct 2024cs.LG cs.AI cs.CVcited by 0

On the map: LLM safety and security → Machine unlearning

Abstract (arXiv)

Gradient-based interpretations often require an anchor point of comparison to avoid saturation in computing feature importance. We show that current baselines defined using static functions--constant mapping, averaging or blurring--inject harmful colour, texture or frequency assumptions that deviate from model behaviour. This leads to accumulation of irregular gradients, resulting in attribution maps that are biased, fragile and manipulable. Departing from the static approach, we propose UNI to compute an (un)learnable, debiased and adaptive baseline by perturbing the input towards an unlearning direction of steepest ascent. Our method discovers reliable baselines and succeeds in erasing salient features, which in turn locally smooths the high-curvature decision boundaries. Our analyses point to unlearning as a promising avenue for generating faithful, efficient and robust interpretations.

Key facts. No citing papers in the map yet (data through 7 Oct 2026). Placed in Machine unlearning, one of 12 directions in LLM safety and security. Builds on 82 of its 109 references that are in the map. First posted on arXiv 10 Oct 2024.

Builds on

82 of 109 references.

Closest ancestors · 3

Same direction, newest first

  1. 2023Fast Model Debias with Machine Unlearning19 Oct
  2. 2022A Survey of Machine Unlearning6 Sep
  3. 2021Unrolling SGD: Understanding Factors Influencing Machine Unlearning27 Sep

Other references · 79

  1. 2015Deep Residual Learning for Image Recognitioncited by 51,809
  2. 2020An Image is Worth 16x16 Words: Transformers for Image Recognition at Scalecited by 29,915
  3. 2019PyTorch: An Imperative Style, High-Performance Deep Learning Librarycited by 20,160
  4. 2021High-Resolution Image Synthesis with Latent Diffusion Modelscited by 18,694
  5. 2023GPT-4 Technical Reportcited by 18,132
  6. 2014Very Deep Convolutional Networks for Large-Scale Image Recognitioncited by 15,267
  7. 2024The Llama 3 Herd of Modelscited by 14,192
  8. 2021Swin Transformer: Hierarchical Vision Transformer using Shifted Windowscited by 12,082
  9. 2018The Unreasonable Effectiveness of Deep Features as a Perceptual Metriccited by 11,029
  10. 2023Visual Instruction Tuningcited by 7,792
  11. 2014Explaining and Harnessing Adversarial Examplescited by 7,287
  12. 2016Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localizationcited by 6,690
38 more
  1. 2016"Why Should I Trust You?": Explaining the Predictions of Any Classifiercited by 5,040
  2. 2022A ConvNet for the 2020scited by 3,545
  3. 2017Axiomatic Attribution for Deep Networkscited by 3,350
  4. 2023Mistral 7Bcited by 3,065
  5. 2019Benchmarking Neural Network Robustness to Common Corruptions and Perturbationscited by 2,857
  6. 2022Locating and Editing Factual Associations in GPTcited by 2,609
  7. 2013Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Mapscited by 2,348
  8. 2013Visualizing and Understanding Convolutional Networkscited by 2,278
  9. 2023Pythia: A Suite for Analyzing Large Language Models Across Training and Scalingcited by 1,734
  10. 2017Towards A Rigorous Science of Interpretable Machine Learningcited by 1,400
  11. 2023Sparse Autoencoders Find Highly Interpretable Features in Language Modelscited by 1,359
  12. 2017Interpretability Beyond Feature Attribution: Quantitative Testing with Concept Activation Vectors (TCAV)cited by 1,309
  13. 2018Sanity Checks for Saliency Mapscited by 1,149
  14. 2022Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 smallcited by 1,130
  15. 2021EfficientNetV2: Smaller Models and Faster Trainingcited by 1,127
  16. 2017SmoothGrad: removing noise by adding noisecited by 1,097
  17. 2020Concept Bottleneck Modelscited by 1,022
  18. 2019Machine Unlearningcited by 1,020
  19. 2014Striving for Simplicity: The All Convolutional Netcited by 983
  20. 2018RISE: Randomized Input Sampling for Explanation of Black-box Modelscited by 825
  21. 2018This Looks Like That: Deep Learning for Interpretable Image Recognitioncited by 731
  22. 2017Network Dissection: Quantifying Interpretability of Deep Visual Representationscited by 703
  23. 2023Towards Automated Circuit Discovery for Mechanistic Interpretabilitycited by 700
  24. 2019Eternal Sunshine of the Spotless Net: Selective Forgetting in Deep Networkscited by 599
  25. 2019Certified Data Removal from Machine Learning Modelscited by 512
  26. 2018Towards Robust Interpretability with Self-Explaining Neural Networkscited by 505
  27. 2018A Benchmark for Interpretability Methods in Deep Neural Networkscited by 467
  28. 2019Making AI Forget You: Data Deletion in Machine Learningcited by 447
  29. 2017Interpretation of Neural Networks is Fragilecited by 443
  30. 2024Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Modelscited by 412
  31. 2016Top-down Neural Attention by Excitation Backpropcited by 318
  32. 2019On the (In)fidelity and Sensitivity for Explanationscited by 309
  33. 2016Not Just a Black Box: Learning Important Features Through Propagating Activation Differencescited by 295
  34. 2017Right for the Right Reasons: Training Differentiable Models by Constraining their Explanationscited by 293
  35. 2019Approximating CNNs with Bag-of-local-Features models works surprisingly well on ImageNetcited by 260
  36. 2018Learning Not to Learn: Training Deep Neural Networks with Biased Datacited by 255
  37. 2019On Completeness-aware Concept-Based Explanations in Deep Neural Networkscited by 239
  38. 2019Explanations can be manipulated and geometry is to blamecited by 215

50 most cited of 79 All 79 (JSON)

Built on by

Not cited yet.