mlsanityData to 14 Oct 2026

B-cos Alignment for Inherently Interpretable CNNs and Vision Transformers

arXiv 2306.10898 ↗First posted 19 Jun 2023cs.CVcited by 31

Not on the map. Why?

Abstract (arXiv)

We present a new direction for increasing the interpretability of deep neural networks (DNNs) by promoting weight-input alignment during training. For this, we propose to replace the linear transformations in DNNs by our novel B-cos transformation. As we show, a sequence (network) of such transformations induces a single linear transformation that faithfully summarises the full model computations. Moreover, the B-cos transformation is designed such that the weights align with relevant signals during optimisation. As a result, those induced linear transformations become highly interpretable and highlight task-relevant features. Importantly, the B-cos transformation is designed to be compatible with existing architectures and we show that it can easily be integrated into virtually all of the latest state of the art models for computer vision - e.g. ResNets, DenseNets, ConvNext models, as well as Vision Transformers - by combining the B-cos-based explanations with normalisation and attention layers, all whilst maintaining similar accuracy on ImageNet. Finally, we show that the resulting explanations are of high visual quality and perform well under quantitative interpretability metrics.

Key facts. Cited by 31 papers in the map (data through 14 Oct 2026). Builds on 60 of its 75 references that are in the map. First posted on arXiv 19 Jun 2023.

Builds on

60 of 75 references.

References · 60

  1. 2017Attention Is All You Needcited by 60,361
  2. 2015Deep Residual Learning for Image Recognitioncited by 52,096
  3. 2014Adam: A Method for Stochastic Optimizationcited by 37,240
  4. 2020An Image is Worth 16x16 Words: Transformers for Image Recognition at Scalecited by 30,186
  5. 2019PyTorch: An Imperative Style, High-Performance Deep Learning Librarycited by 20,305
  6. 2014Very Deep Convolutional Networks for Large-Scale Image Recognitioncited by 15,364
  7. 2015Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shiftcited by 7,005
  8. 2016Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localizationcited by 6,733
  9. 2016Densely Connected Convolutional Networkscited by 6,720
  10. 2017A Unified Approach to Interpreting Model Predictionscited by 6,555
  11. 2016Layer Normalizationcited by 5,669
  12. 2017mixup: Beyond Empirical Risk Minimizationcited by 5,183
38 more
  1. 2016"Why Should I Trust You?": Explaining the Predictions of Any Classifiercited by 5,070
  2. 2022A ConvNet for the 2020scited by 3,579
  3. 2017Axiomatic Attribution for Deep Networkscited by 3,384
  4. 2019CutMix: Regularization Strategy to Train Strong Classifiers with Localizable Featurescited by 2,884
  5. 2016Aggregated Residual Transformations for Deep Neural Networkscited by 2,546
  6. 2015Learning Deep Features for Discriminative Localizationcited by 2,436
  7. 2013Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Mapscited by 2,371
  8. 2017Improved Regularization of Convolutional Neural Networks with Cutoutcited by 1,628
  9. 2018Group Normalizationcited by 1,524
  10. 2017Learning Important Features Through Propagating Activation Differencescited by 1,458
  11. 2017Interpretability Beyond Feature Attribution: Quantitative Testing with Concept Activation Vectors (TCAV)cited by 1,319
  12. 2017Random Erasing Data Augmentationcited by 1,239
  13. 2016Instance Normalization: The Missing Ingredient for Fast Stylizationcited by 1,198
  14. 2018Sanity Checks for Saliency Mapscited by 1,156
  15. 2014Striving for Simplicity: The All Convolutional Netcited by 990
  16. 2018Robustness May Be at Odds with Accuracycited by 880
  17. 2018RISE: Randomized Input Sampling for Explanation of Black-box Modelscited by 838
  18. 2020Quantifying Attention Flow in Transformerscited by 775
  19. 2018This Looks Like That: Deep Learning for Interpretable Image Recognitioncited by 737
  20. 2020Transformer Interpretability Beyond Attention Visualizationcited by 627
  21. 2021Early Convolutions Help Transformers See Bettercited by 396
  22. 2019Is Attention Interpretable?cited by 389
  23. 2014On the Number of Linear Regions of Deep Neural Networkscited by 381
  24. 2021High-Performance Large-Scale Image Recognition Without Normalizationcited by 360
  25. 2019Approximating CNNs with Bag-of-local-Features models works surprisingly well on ImageNetcited by 260
  26. 2021TrivialAugment: Tuning-free Yet State-of-the-Art Data Augmentationcited by 227
  27. 2019Full-Gradient Representation for Neural Network Visualizationcited by 191
  28. 2013Maxout Networkscited by 187
  29. 2020The elephant in the interpretability room: Why use attention as explanation when we have saliency methods?cited by 122
  30. 2022Better plain ViT baselines for ImageNet-1kcited by 98
  31. 2019Robust and interpretable blind image denoising via bias-free convolutional neural networkscited by 90
  32. 2019Benchmarking Attribution Methods with Relative Feature Importancecited by 89
  33. 2022B-cos Networks: Alignment is All We Need for Interpretabilitycited by 88
  34. 2019MultiGrain: a unified image embedding for classes and instancescited by 85
  35. 2017Learning how to explain neural networks: PatternNet and PatternAttributioncited by 84
  36. 2017Cosine Normalization: Using Cosine Similarity Instead of Dot Product in Neural Networkscited by 59
  37. 2017Deep Hyperspherical Learningcited by 55
  38. 2021Convolutional Dynamic Alignment Networks for Interpretable Classificationscited by 54

50 most cited of 60 All 60 (JSON)

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2023: 1; 2024: 9; 2025: 13; 2026: 8.

31 papers.

Sort:

By area:

  1. 2023Frozen Transformers in Language Models Are Effective Visual Encoder Layerscited by 39
  2. 2024A Comprehensive Survey for Hyperspectral Image Classification: The Evolution from Conventional to Transformers and Mamba Modelscited by 19
  3. 2025A Comprehensive Survey on Self-Interpretable Neural Networkscited by 18
  4. 2024Good Teachers Explain: Explanation-Enhanced Knowledge Distillationcited by 14
  5. 2024B-cosification: Transforming Deep Neural Networks to be Inherently Interpretablecited by 12
  6. 2025CoE: Chain-of-Explanation via Automatic Visual Concept Circuit Description and Polysemanticity Quantificationcited by 9
  7. 2024Benchmarking the Attribution Quality of Vision Modelscited by 8
  8. 2024Benchmarking XAI Explanations with Human-Aligned Evaluationscited by 8
  9. 2024Classification Metrics for Image Explanations: Towards Building Reliable XAI-Evaluationscited by 7
  10. 2025Explaining Similarity in Vision-Language Encoders with Weighted Banzhaf Interactionscited by 6
  11. 2025B-cos LM: Efficiently Transforming Pre-trained Language Models for Improved Explainabilitycited by 4
  12. 2026DAVE: Distribution-aware Attribution via ViT Gradient Decompositioncited by 4
19 more
  1. 2025Birds look like cars: Adversarial analysis of intrinsically interpretable deep learningcited by 3
  2. 2024OMENN: One Matrix to Explain Neural Networkscited by 2
  3. 2025From Pixels to Perception: Interpretable Predictions via Instance-wise Grouped Feature Selectioncited by 2
  4. 2024Attri-Net: A Globally and Locally Inherently Interpretable Model for Multi-Label Classification Using Class-Specific Counterfactualscited by 1
  5. 2025COMIX: Compositional Explanations using Prototypescited by 1
  6. 2025Disentangling Visual Transformers: Patch-level Interpretability for Image Classificationcited by 1
  7. 2025FaCT: Faithful Concept Traces for Explaining Neural Network Decisionscited by 1
  8. 2026Exploring SAIG Methods for an Objective Evaluation of XAIcited by 1
  9. 2026End-to-end Feature Alignment: A Simple CNN with Intrinsic Class Attributioncited by 1
  10. 2024IMAFD: An Interpretable Multi-stage Approach to Flood Detection from time series Multispectral Datacited by 0
  11. 2025SIC: Similarity-Based Interpretable Image Classification with Neural Networkscited by 0
  12. 2025PiPViT: Patch-based Visual Interpretable Prototypes for Retinal Image Analysiscited by 0
  13. 2025Interpretable Diffusion Models with B-cos Networkscited by 0
  14. 2025Pulling Back the Curtain on Deep Networkscited by 0
  15. 2026DINO-QPM: Adapting Visual Foundation Models for Globally Interpretable Image Classificationcited by 0
  16. 2026When Can We Trust Deep Neural Networks? Towards Reliable Industrial Deployment with an Interpretability Guidecited by 0
  17. 2026B-cos GNNs: Faithful Explanations through Dynamic Linearitycited by 0
  18. 2026GRAPE: Graph-Augmented Prototype Explanations for Interactive Medical Image Diagnosiscited by 0
  19. 2026qZACH-ViT: Quantization-Aware Intrinsic Explanations with Recursive Attribution-Stabilized Optimizationcited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2023BDC-Adapter: Brownian Distance Covariance for Better Vision-Language Reasoning
  2. 2026Sparse but not Simpler: A Multi-Level Interpretability Analysis of Vision Transformers
  3. 2026Linearizing Vision Transformer with Test-Time Training
  4. 2026Cognitive Alignment At No Cost: Inducing Human Attention Biases For Interpretable Vision Transformers
  5. 2025Prompt the Unseen: Evaluating Visual-Language Alignment Beyond Supervision
  6. 2025ViT-Linearizer: Distilling Quadratic Knowledge into Linear-Time Vision Models
  7. 2026Can Cross-Layer Transcoders Replace Vision Transformer Activations? An Interpretable Perspective on Vision
  8. 2022The SVD of Convolutional Weights: A CNN Interpretability Framework