mlsanityData to 14 Oct 2026

What do Vision Transformers Learn? A Visual Exploration

arXiv 2212.06727 ↗First posted 13 Dec 2022cs.CVcited by 76

Not on the map. Why?

Abstract (arXiv)

Vision transformers (ViTs) are quickly becoming the de-facto architecture for computer vision, yet we understand very little about why they work and what they learn. While existing studies visually analyze the mechanisms of convolutional neural networks, an analogous exploration of ViTs remains challenging. In this paper, we first address the obstacles to performing visualizations on ViTs. Assisted by these solutions, we observe that neurons in ViTs trained with language model supervision (e.g., CLIP) are activated by semantic concepts rather than visual features. We also explore the underlying differences between ViTs and CNNs, and we find that transformers detect image background features, just like their convolutional counterparts, but their predictions depend far less on high-frequency information. On the other hand, both architecture types behave similarly in the way features progress from abstract patterns in early layers to concrete objects in late layers. In addition, we show that ViTs maintain spatial information in all layers except the final layer. In contrast to previous works, we show that the last layer most likely discards the spatial information and behaves as a learned global pooling operation. Finally, we conduct large-scale visualizations on a wide range of ViT variants, including DeiT, CoaT, ConViT, PiT, Swin, and Twin, to validate the effectiveness of our method.

Key facts. Cited by 76 papers in the map (data through 14 Oct 2026). Builds on 27 of its 37 references that are in the map. First posted on arXiv 13 Dec 2022.

Builds on

27 of 37 references.

References · 27

  1. 2021Learning Transferable Visual Models From Natural Language Supervisioncited by 33,744
  2. 2020An Image is Worth 16x16 Words: Transformers for Image Recognition at Scalecited by 30,186
  3. 2021Masked Autoencoders Are Scalable Vision Learnerscited by 7,403
  4. 2020Training data-efficient image transformers & distillation through attentioncited by 4,586
  5. 2013Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Mapscited by 2,371
  6. 2013Visualizing and Understanding Convolutional Networkscited by 2,290
  7. 2018ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustnesscited by 1,432
  8. 2019Certified Adversarial Robustness via Randomized Smoothingcited by 1,340
  9. 2021Swin Transformer V2: Scaling Up Capacity and Resolutioncited by 1,228
  10. 2017SmoothGrad: removing noise by adding noisecited by 1,109
  11. 2021Scaling Vision Transformerscited by 999
  12. 2021Do Vision Transformers See Like Convolutional Neural Networks?cited by 675
15 more
  1. 2020Transformer Interpretability Beyond Attention Visualizationcited by 627
  2. 2021CoAtNet: Marrying Convolution and Attention for All Data Sizescited by 569
  3. 2021Twins: Revisiting the Design of Spatial Attention in Vision Transformerscited by 565
  4. 2021ConViT: Improving Vision Transformers with Soft Convolutional Inductive Biasescited by 484
  5. 2021Intriguing Properties of Vision Transformerscited by 446
  6. 2019Dreaming to Distill: Data-free Knowledge Transfer via DeepInversioncited by 423
  7. 2014Understanding Deep Image Representations by Inverting Themcited by 403
  8. 2021Rethinking Spatial Dimensions of Vision Transformerscited by 361
  9. 2022How Do Vision Transformers Work?cited by 307
  10. 2021Co-Scale Conv-Attentional Image Transformerscited by 224
  11. 2021Vision Transformers are Robust Learnerscited by 207
  12. 2021PeCo: Perceptual Codebook for BERT Pre-training of Vision Transformerscited by 204
  13. 2015Inverting Visual Representations with Convolutional Networkscited by 170
  14. 2021How Well do Feature Visualizations Support Causal Understanding of CNN Activations?cited by 28
  15. 2022Plug-In Inversion: Model-Agnostic Inversion for Vision with Data Augmentationscited by 18

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2023: 22; 2024: 24; 2025: 15; 2026: 15.

76 papers.

Sort:

By area:

  1. 2023Segment Anything in High Qualitycited by 376
  2. 2023A Cookbook of Self-Supervised Learningcited by 250
  3. 2024LHRS-Bot: Empowering Remote Sensing with VGI-Enhanced Large Multimodal Language Modelcited by 138
  4. 2023Battle of the Backbones: A Large-Scale Comparison of Pretrained Models across Computer Vision Taskscited by 72
  5. 2023Cinematic Mindscapes: High-quality Video Reconstruction from Brain Activitycited by 71
  6. 2024BoQ: A Place is Worth a Bag of Learnable Queriescited by 64
  7. 2025Universal Sparse Autoencoders: Interpretable Cross-Model Concept Alignmentcited by 62
  8. 2023AttentionViz: A Global View of Transformer Attentioncited by 50
  9. 2024IRSAM: Advancing Segment Anything Model for Infrared Small Target Detectioncited by 46
  10. 2023Channel Vision Transformers: An Image Is Worth 1 x 16 x 16 Wordscited by 34
  11. 2023ViT-Lens: Towards Omni-modal Representationscited by 29
  12. 2023Unlocking Feature Visualization for Deeper Networks with MAgnitude Constrained Optimizationcited by 27
38 more
  1. 2024LiFT: A Surprisingly Simple Lightweight Feature Transform for Dense ViT Descriptorscited by 27
  2. 2023Analyzing Vision Transformers for Image Classification in Class Embedding Spacecited by 26
  3. 2025Modality-Aware Neuron Pruning for Unlearning in Multimodal Large Language Modelscited by 26
  4. 2025Interpretable and Testable Vision Features via Sparse Autoencoderscited by 25
  5. 2024Understanding Video Transformers via Universal Concept Discoverycited by 23
  6. 2023Multimodal Pretraining of Medical Time Series and Notescited by 22
  7. 2023TagCLIP: A Local-to-Global Framework to Enhance Open-Vocabulary Multi-Label Classification of CLIP Without Trainingcited by 19
  8. 2024Decomposing and Editing Predictions by Modeling Model Computationcited by 19
  9. 2023Scale Alone Does not Improve Mechanistic Interpretability in Vision Modelscited by 18
  10. 2024Targeted Unlearning with Single Layer Unlearning Gradientcited by 18
  11. 2024Towards Generalizable Scene Change Detectioncited by 17
  12. 2023Self-attention in Vision Transformers Performs Perceptual Grouping, Not Attentioncited by 16
  13. 2023Out-of-Distribution Detection by Leveraging Between-Layer Transformation Smoothnesscited by 15
  14. 2023Sketch-A-Shape: Zero-Shot Sketch-to-3D Shape Generationcited by 14
  15. 2024If CLIP Could Talk: Understanding Vision-Language Model Representations Through Their Preferred Concept Descriptionscited by 14
  16. 2024What do we learn from inverting CLIP models?cited by 13
  17. 2024TextRefiner: Internal Visual Feature as Efficient Refiner for Vision-Language Models Prompt Tuningcited by 13
  18. 2023Explainability of Vision Transformers: A Comprehensive Review and New Perspectivescited by 12
  19. 2024EIVEN: Efficient Implicit Attribute Value Extraction using Multimodal LLMcited by 12
  20. 2024Dissecting Query-Key Interaction in Vision Transformerscited by 12
  21. 2025ToSA: Token Merging with Spatial Awarenesscited by 11
  22. 2023Contextual Vision Transformers for Robust Representation Learningcited by 10
  23. 2025Mettle: Meta-Token Learning for Memory-Efficient Audio-Visual Adaptationcited by 10
  24. 2023Single Cells Are Spatial Tokens: Transformers for Spatial Transcriptomic Data Imputationcited by 7
  25. 2024Instruction-Guided Fusion of Multi-Layer Visual Features in Large Vision-Language Modelscited by 7
  26. 2025Discovering Influential Neuron Path in Vision Transformerscited by 7
  27. 2024Causal Graphical Models for Vision-Language Compositional Understandingcited by 6
  28. 2025Unlocking Pretrained Vision Transformers for Time Series Classificationcited by 6
  29. 2024Feature Accentuation: Revealing 'What' Features Respond to in Natural Imagescited by 5
  30. 2026Attentive multilayer fusion for vision transformerscited by 4
  31. 2026Steerable Visual Representationscited by 4
  32. 2026Investigating Single-Block Recurrence in Vision Transformers for Image Recognitioncited by 4
  33. 2025Evaluating Sample Utility for Efficient Data Selection by Mimicking Model Weightscited by 3
  34. 2025LLMs Can Compensate for Deficiencies in Visual Representationscited by 3
  35. 2025Boosting Generative Adversarial Transferability with Self-supervised Vision Transformer Featurescited by 3
  36. 2025Self-Supervised Visual Prompting for Cross-Domain Road Damage Detectioncited by 3
  37. 2026A Single Neuron Is Sufficient to Bypass Safety Alignment in Large Language Modelscited by 3
  38. 2023Learning to Generate Training Datasets for Robust Semantic Segmentationcited by 2

50 most cited of 76 All 76 (JSON)