# From Hope to Safety: Unlearning Biases of Deep Models via Gradient Penalization in Latent Space arXiv: https://arxiv.org/abs/2308.09437 · HTML: https://mlsanity.com/paper/2308.09437/ ## Key facts - Cited by 20 papers in the map (data through 14 Oct 2026). - Placed in Machine unlearning, one of 12 directions in LLM safety and security. - Builds on 19 of its 34 references that are in the map. - First posted on arXiv 18 Aug 2023. ## Placed in - LLM safety and security / [Machine unlearning](https://mlsanity.com/safety/machine-unlearning/) ## Builds on (most cited first, 19 of 19) - [Deep Residual Learning for Image Recognition](https://mlsanity.com/paper/1512.03385/) (2015) - [Very Deep Convolutional Networks for Large-Scale Image Recognition](https://mlsanity.com/paper/1409.1556/) (2014) - [EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks](https://mlsanity.com/paper/1905.11946/) (2019) - [Deep Learning Face Attributes in the Wild](https://mlsanity.com/paper/1411.7766/) (2014) - [Visualizing and Understanding Convolutional Networks](https://mlsanity.com/paper/1311.2901/) (2013) - [Interpretability Beyond Feature Attribution: Quantitative Testing with Concept Activation Vectors (TCAV)](https://mlsanity.com/paper/1711.11279/) (2017) - [Skin Lesion Analysis Toward Melanoma Detection: A Challenge at the 2017 International Symposium on Biomedical Imaging (ISBI), Hosted by the International Skin Imaging Collaboration (ISIC)](https://mlsanity.com/paper/1710.05006/) (2017) - [Unmasking Clever Hans Predictors and Assessing What Machines Really Learn](https://mlsanity.com/paper/1902.10178/) (2019) - [Right for the Right Reasons: Training Differentiable Models by Constraining their Explanations](https://mlsanity.com/paper/1703.03717/) (2017) - [BCN20000: Dermoscopic Lesions in the Wild](https://mlsanity.com/paper/1908.02288/) (2019) - [Interpretations are useful: penalizing explanations to align neural networks with prior knowledge](https://mlsanity.com/paper/1909.13584/) (2019) - [Making deep neural networks right for the right scientific reasons by interacting with their explanations](https://mlsanity.com/paper/2001.05371/) (2020) - [Can contrastive learning avoid shortcut solutions?](https://mlsanity.com/paper/2106.11230/) (2021) - [A Whac-A-Mole Dilemma: Shortcuts Come in Multiples Where Mitigating One Amplifies Others](https://mlsanity.com/paper/2212.04825/) (2022) - [Towards Best Practice in Explaining Neural Network Decisions with LRP](https://mlsanity.com/paper/1910.09840/) (2019) - [Causally motivated Shortcut Removal Using Auxiliary Labels](https://mlsanity.com/paper/2105.06422/) (2021) - [Concept-level Debugging of Part-Prototype Networks](https://mlsanity.com/paper/2205.15769/) (2022) - [Finding and Fixing Spurious Patterns with Explanations](https://mlsanity.com/paper/2106.02112/) (2021) - [Robust Semantic Interpretability: Revisiting Concept Activation Vectors](https://mlsanity.com/paper/2104.02768/) (2021) ## Built on by (most cited first, 20 of 20) - [Rethinking Machine Unlearning for Large Language Models](https://mlsanity.com/paper/2402.08787/) (2024) - [Concept-Based Explanations in Computer Vision: Where Are We and Where Could We Go?](https://mlsanity.com/paper/2409.13456/) (2024) - [Contestable AI needs Computational Argumentation](https://mlsanity.com/paper/2405.10729/) (2024) - [Mitigating Clever Hans Strategies in Image Classifiers through Generating Counterexamples](https://mlsanity.com/paper/2510.17524/) (2025) - [Controllable Continual Test-Time Adaptation](https://mlsanity.com/paper/2405.14602/) (2024) - [Reactive Model Correction: Mitigating Harm to Task-Relevant Features via Conditional Bias Suppression](https://mlsanity.com/paper/2404.09601/) (2024) - [Visual Disentangled Diffusion Autoencoders: Scalable Counterfactual Generation for Foundation Models](https://mlsanity.com/paper/2601.21851/) (2026) - [Attribution-Guided Pruning for Insight and Control: Circuit Discovery and Targeted Correction in Small-scale LLMs](https://mlsanity.com/paper/2506.13727/) (2025) - [On The Variability of Concept Activation Vectors](https://mlsanity.com/paper/2509.24058/) (2025) - [Contrastive Sparse Autoencoders for Interpreting Planning of Chess-Playing Agents](https://mlsanity.com/paper/2406.04028/) (2024) - [Towards certifiable AI in aviation: landscape, challenges, and opportunities](https://mlsanity.com/paper/2409.08666/) (2024) - [Ensuring Medical AI Safety: Interpretability-Driven Detection and Mitigation of Spurious Model Behavior and Associated Data](https://mlsanity.com/paper/2501.13818/) (2025) - [Concept activation vectors: a unifying view and adversarial attacks](https://mlsanity.com/paper/2509.22755/) (2025) - [TDHook: A Lightweight Framework for Interpretability](https://mlsanity.com/paper/2509.25475/) (2025) - [Probing the Probes: Methods and Metrics for Concept Alignment](https://mlsanity.com/paper/2511.04312/) (2025) - [E-TCAV: Formalizing Penultimate Proxies for Efficient Concept Based Interpretability](https://mlsanity.com/paper/2605.10261/) (2026) - [α-TCAV: A Unified Framework for Testing with Concept Activation Vectors](https://mlsanity.com/paper/2605.15688/) (2026) - [Bias Leaves a Gradient Trail: Label-Free Bias Identification via Gradient Probes on Concept Decompositions](https://mlsanity.com/paper/2605.28780/) (2026) - [ICON decomposition: Multivariate concept auditing reveals shortcuts that neural networks rely on](https://mlsanity.com/paper/2608.26083/) (2026) - [Towards Fast and Disentangled Counterfactuals for Visual Foundation Models](https://mlsanity.com/paper/2610.00895/) (2026) ## Abstract (arXiv) Deep Neural Networks are prone to learning spurious correlations embedded in the training data, leading to potentially biased predictions. This poses risks when deploying these models for high-stake decision-making, such as in medical applications. Current methods for post-hoc model correction either require input-level annotations which are only possible for spatially localized biases, or augment the latent feature space, thereby hoping to enforce the right reasons. We present a novel method for model correction on the concept level that explicitly reduces model sensitivity towards biases via gradient penalization. When modeling biases via Concept Activation Vectors, we highlight the importance of choosing robust directions, as traditional regression-based approaches such as Support Vector Machines tend to result in diverging directions. We effectively mitigate biases in controlled and real-world settings on the ISIC, Bone Age, ImageNet and CelebA datasets using VGG, ResNet and EfficientNet architectures. Code is available on https://github.com/frederikpahde/rrclarc.