mlsanityData to 14 Oct 2026

The Tunnel Effect: Building Data Representations in Deep Neural Networks

arXiv 2305.19753 ↗First posted 31 May 2023cs.LG cs.CVcited by 34

Not on the map. Why?

Abstract (arXiv)

Deep neural networks are widely known for their remarkable effectiveness across various tasks, with the consensus that deeper networks implicitly learn more complex data representations. This paper shows that sufficiently deep networks trained for supervised image classification split into two distinct parts that contribute to the resulting data representations differently. The initial layers create linearly-separable representations, while the subsequent layers, which we refer to as textitthe tunnel, compress these representations and have a minimal impact on the overall performance. We explore the tunnel's behavior through comprehensive empirical studies, highlighting that it emerges early in the training process. Its depth depends on the relation between the network's capacity and task complexity. Furthermore, we show that the tunnel degrades out-of-distribution generalization and discuss its implications for continual learning.

Key facts. Cited by 34 papers in the map (data through 14 Oct 2026). Builds on 54 of its 72 references that are in the map. First posted on arXiv 31 May 2023.

Builds on

54 of 72 references.

References · 54

  1. 2015Deep Residual Learning for Image Recognitioncited by 52,096
  2. 2014Very Deep Convolutional Networks for Large-Scale Image Recognitioncited by 15,364
  3. 2018UMAP: Uniform Manifold Approximation and Projection for Dimension Reductioncited by 2,898
  4. 2018Neural Tangent Kernel: Convergence and Generalization in Neural Networkscited by 2,621
  5. 2019Similarity of Neural Network Representations Revisitedcited by 1,755
  6. 2014How transferable are features in deep neural networks?cited by 1,225
  7. 2018Robustness May Be at Odds with Accuracycited by 880
  8. 2020Prevalence of Neural Collapse during the terminal phase of deep learning trainingcited by 719
  9. 2018Universal Transformerscited by 647
  10. 2019Deep Equilibrium Modelscited by 615
  11. 2018Rethinking the Value of Network Pruningcited by 599
  12. 2017Opening the Black Box of Deep Neural Networks via Informationcited by 503
38 more
  1. 2014On the Number of Linear Regions of Deep Neural Networkscited by 381
  2. 2020What is being transferred in transfer learning?cited by 368
  3. 2016On the Expressive Power of Deep Neural Networkscited by 306
  4. 2019Intrinsic dimension of data representations in deep neural networkscited by 282
  5. 2020Do Wide and Deep Networks Learn the Same Things? Uncovering How Neural Network Representations Vary with Width and Depthcited by 245
  6. 2021A Geometric Analysis of Neural Collapse with Unconstrained Featurescited by 227
  7. 2020The large learning rate phase of deep learning: the catapult mechanismcited by 216
  8. 2018Deep learning generalizes because the parameter-function map is biased towards simple functionscited by 183
  9. 2019Complexity of Linear Regions in Deep Networkscited by 167
  10. 2020Anatomy of Catastrophic Forgetting: Hidden Representations and Task Semanticscited by 161
  11. 2021The Low-Rank Simplicity Bias in Deep Networkscited by 154
  12. 2020Adversarial Continual Learningcited by 142
  13. 2018Regularization Matters: Generalization and Optimization of Neural Nets v.s. their Induced Kernelcited by 131
  14. 2023Similarity of Neural Network Models: A Survey of Functional and Representational Measurescited by 131
  15. 2021Generalized Shape Metrics on Neural Representationscited by 112
  16. 2020What makes instance discrimination good for transfer learning?cited by 111
  17. 2019One ticket to win them all: generalizing lottery ticket initializations across datasets and optimizerscited by 102
  18. 2021On the Role of Neural Collapse in Transfer Learningcited by 100
  19. 2019Latent Replay for Real-Time Continual Learningcited by 97
  20. 2019Are All Layers Created Equal?cited by 94
  21. 2021Continual Learning in the Teacher-Student Setup: Impact of Task Similaritycited by 88
  22. 2015Representation Benefits of Deep Feedforward Networkscited by 86
  23. 2020Towards Learning a Universal Non-Semantic Representation of Speechcited by 84
  24. 2022Head2Toe: Utilizing Intermediate Representations for Better Transfer Learningcited by 83
  25. 2022How catastrophic can catastrophic forgetting be in linear regression?cited by 81
  26. 2021A Broad Study on the Transferability of Visual Representations with Contrastive Learningcited by 78
  27. 2019A Fine-Grained Spectral Perspective on Neural Networkscited by 68
  28. 2021Wide Neural Networks Forget Less Catastrophicallycited by 64
  29. 2022Limitations of Neural Collapse for Understanding Generalization in Deep Learningcited by 63
  30. 2019Universality and individuality in neural dynamics across large populations of recurrent networkscited by 61
  31. 2018Trellis Networks for Sequence Modelingcited by 52
  32. 2022Architecture Matters in Continual Learningcited by 46
  33. 2018Recurrent Stacking of Layers for Compact Neural Machine Translation Modelscited by 38
  34. 2017Transfer learning for music classification and regression taskscited by 37
  35. 2021Revisiting the Transferability of Supervised Pretraining: an MLP Perspectivecited by 36
  36. 2021Fast Adaptation with Linearized Neural Networkscited by 32
  37. 2022Understanding and Improving Transfer Learning of Deep Models via Neural Collapsecited by 24
  38. 2021Feature Learning and Signal Propagation in Deep Neural Networkscited by 18

50 most cited of 54 All 54 (JSON)

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2024: 14; 2025: 14; 2026: 6.

34 papers.

Sort:

By area:

  1. 2024Linguistic Collapse: Neural Collapse in (Large) Language Modelscited by 38
  2. 2024Parseval Regularization for Continual Reinforcement Learningcited by 21
  3. 2024Neural Collapse versus Low-rank Bias: Is Deep Neural Collapse Really Optimal?cited by 17
  4. 2025Intermediate Layer Classifiers for OOD generalizationcited by 14
  5. 2024What Variables Affect Out-of-Distribution Generalization in Pretrained Models?cited by 13
  6. 2024A Global Geometric Analysis of Maximal Coding Rate Reductioncited by 13
  7. 2025Parameter Symmetry Potentially Unifies Deep Learning Theorycited by 12
  8. 2025Controlling Neural Collapse Enhances Out-of-Distribution Detection and Transfer Learningcited by 11
  9. 2024Formation of Representations in Neural Networkscited by 10
  10. 2025Neural Collapse is Globally Optimal in Deep Regularized ResNets and Transformerscited by 10
  11. 2025Register and [CLS] tokens yield a decoupling of local and global features in large ViTscited by 9
  12. 2024End-to-End Training Induces Information Bottleneck through Layer-Role Differentiation: A Comparative Analysis with Layer-wise Trainingcited by 8
22 more
  1. 2024Neural Collapse Meets Differential Privacy: Curious Behaviors of NoisyGD with Near-perfect Representation Learningcited by 8
  2. 2025Proof of a perfect platonic representation hypothesiscited by 8
  3. 2024Effective Layer Pruning Through Similarity Metric Perspectivecited by 6
  4. 2025Neural Collapse in Cumulative Link Models for Ordinal Regression: An Analysis with Unconstrained Feature Modelcited by 6
  5. 2025Why and How Auxiliary Tasks Improve JEPA Representationscited by 5
  6. 2025Efficient LLM Moderation with Multi-Layer Latent Prototypescited by 4
  7. 2025Synthetic Speech Source Tracing using Metric Learningcited by 4
  8. 2024Towards Optimal Adapter Placement for Efficient Transfer Learningcited by 3
  9. 2025Unpacking Softmax: How Temperature Drives Representation Collapse, Compression, and Generalizationcited by 3
  10. 2024Improving Continual Learning Performance and Efficiency with Auxiliary Classifierscited by 2
  11. 2024Quantifying uncertainty in lung cancer segmentation with foundation models applied to mixed-domain datasetscited by 2
  12. 2024DualFed: Enjoying both Generalization and Personalization in Federated Learning via Hierachical Representationscited by 2
  13. 2025Linearly Separable Features in Shallow Nonlinear Networks: Width Scales Polynomially with Intrinsic Data Dimensioncited by 2
  14. 2026Diagnosing Generalization Failures from Representational Geometry Markerscited by 2
  15. 2026Optimizer choice matters for the emergence of Neural Collapsecited by 1
  16. 2024Layer Pruning with Consensus: A Triple-Win Solutioncited by 0
  17. 2025Random forest-based out-of-distribution detection for robust lung cancer segmentationcited by 0
  18. 2025Logit-Based Losses Limit the Effectiveness of Feature Knowledge Distillationcited by 0
  19. 2026Sparsity, Superposition, and Forgetting: A Mechanistic Study of Representation Retention in Continual Learningcited by 0
  20. 2026Lifelong Representations: A Survey on Continual Self-Supervised Learning for Vision Modelscited by 0
  21. 2026Simplifying Neural Networks During Trainingcited by 0
  22. 2026The Layer Mystery of VLA: An Information-Theoretical Analysis of VLA Latent Interfacecited by 0