Post-Hoc Concept Disentanglement: From Correlated to Isolated Concept Representations
On the map: Mechanistic interpretability → Sparse autoencoder training and evaluation
Abstract (arXiv)
Concept Activation Vectors (CAVs) are widely used to model human-understandable concepts as directions within the latent space of neural networks. They are trained by identifying directions from the activations of concept samples to those of non-concept samples. However, this method often produces similar, non-orthogonal directions for correlated concepts, such as "beard" and "necktie" within the CelebA dataset, which frequently co-occur in images of men. This entanglement complicates the interpretation of concepts in isolation and can lead to undesired effects in CAV applications, such as activation steering. To address this issue, we introduce a post-hoc concept disentanglement method that employs a non-orthogonality loss, facilitating the identification of orthogonal concept directions while preserving directional correctness. We evaluate our approach with real-world and controlled correlated concepts in CelebA and a synthetic FunnyBirds dataset with VGG16 and ResNet18 architectures. We further demonstrate the superiority of orthogonalized concept representations in activation steering tasks, allowing (1) the insertion of isolated concepts into input images through generative models and (2) the removal of concepts for effective shortcut suppression with reduced impact on correlated concepts in comparison to baseline CAVs.
Key facts. Cited by 1 paper in the map (data through 14 Oct 2026). Placed in Sparse autoencoder training and evaluation, one of 13 directions in Mechanistic interpretability. Builds on 36 of its 50 references that are in the map. First posted on arXiv 7 Mar 2025.
Builds on
36 of 50 references.
Closest ancestors · 1
Same direction, newest first
- 2022Toy Models of Superposition21 Sep
Other references · 35
- 2015Deep Residual Learning for Image Recognitioncited by 52,096
- 2014Very Deep Convolutional Networks for Large-Scale Image Recognitioncited by 15,364
- 2016Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localizationcited by 6,733
- 2017A Unified Approach to Interpreting Model Predictionscited by 6,555
- 2014Deep Learning Face Attributes in the Wildcited by 3,693
- 2017Axiomatic Attribution for Deep Networkscited by 3,384
- 2020Shortcut Learning in Deep Neural Networkscited by 1,969
- 2017Interpretability Beyond Feature Attribution: Quantitative Testing with Concept Activation Vectors (TCAV)cited by 1,319
- 2018Learning deep representations by mutual information estimation and maximizationcited by 1,273
- 2017SmoothGrad: removing noise by adding noisecited by 1,109
- 2014Striving for Simplicity: The All Convolutional Netcited by 990
- 2018Disentangling by Factorisingcited by 616
23 more
- 2020Null It Out: Guarding Protected Attributes by Iterative Nullspace Projectioncited by 455
- 2023Emergent Linear Representations in World Models of Self-Supervised Sequence Modelscited by 319
- 2022Post-hoc Concept Bottleneck Modelscited by 293
- 2020Understanding the Role of Individual Units in a Deep Neural Networkcited by 255
- 2020Concept Whitening for Interpretable Image Recognitioncited by 225
- 2022CRAFT: Concept Recursive Activation FacTorization for Explainabilitycited by 178
- 2017Learning to Generate Reviews and Discovering Sentimentcited by 148
- 2022From Attribution Maps to Human-Understandable Explanations through Concept Relevance Propagationcited by 141
- 2020Invertible Concept-based Explanations for CNN Models with Non-negative Concept Activation Vectorscited by 120
- 2022Disentangled Representation Learningcited by 112
- 2017Multi-Level Variational Autoencoder: Learning Disentangled Representations from Grouped Observationscited by 100
- 2020On Disentangled Representations Learned From Correlated Datacited by 94
- 2018FineGAN: Unsupervised Hierarchical Disentanglement for Fine-Grained Object Generation and Discoverycited by 58
- 2021Software for Dataset-wide XAI: From Local Explanations to Global Insights with Zennit, CoRelAy, and ViRelAycited by 55
- 2023Multi-dimensional concept discovery (MCD): A unifying framework with completeness guaranteescited by 50
- 2023FunnyBirds: A Synthetic Vision Dataset for a Part-Based Analysis of Explainable AI Methodscited by 50
- 2022Probing Classifiers are Unreliable for Concept Removal and Detectioncited by 41
- 2022Spurious Features Everywhere -- Large-Scale Detection of Harmful Spurious Features in ImageNetcited by 41
- 2022Navigating Neural Space: Revisiting Concept Activation Vectors to Overcome Directional Divergencecited by 25
- 2022Disentangled Explanations of Neural Network Predictions by Finding Relevant Subspacescited by 25
- 2023Reveal to Revise: An Explainable AI Life Cycle for Iterative Bias Correction of Deep Modelscited by 17
- 2019Concept Saliency Maps to Visualize Relevant Features in Deep Generative Modelscited by 11
- 2024Reactive Model Correction: Mitigating Harm to Task-Relevant Features via Conditional Bias Suppressioncited by 5
Built on by
Citing papers per year: 2026: 1.
1 paper.
Similar, not cited
Close in wording, no citation either way
- 2026Disentangling Steering Vectors
- 2025Concept-SAE: A Controllable and Invertible Concept Interface for Sparse Autoencoders
- 2023Disentangling Neuron Representations with Concept Vectors
- 2025Toward a Flexible Framework for Linear Representation Hypothesis Using Maximum Likelihood Estimation
- 2025From Isolation to Entanglement: When Do Interpretability Methods Identify and Disentangle Known Concepts?
- 2024OmniPrism: Learning Disentangled Visual Concept for Image Generation
- 2022Exploring Concept Contribution Spatially: Hidden Layer Interpretation with Spatial Activation Concept Vector
- 2026On the Faithfulness of Post-Hoc Concept Bottleneck Models