mlsanityData to 14 Oct 2026

What Do Compressed Deep Neural Networks Forget?

arXiv 1911.05248 ↗First posted 13 Nov 2019cs.LG cs.AI cs.CVcited by 171

Not on the map. Why?

Abstract (arXiv)

Deep neural network pruning and quantization techniques have demonstrated it is possible to achieve high levels of compression with surprisingly little degradation to test set accuracy. However, this measure of performance conceals significant differences in how different classes and images are impacted by model compression techniques. We find that models with radically different numbers of weights have comparable top-line performance metrics but diverge considerably in behavior on a narrow subset of the dataset. This small subset of data points, which we term Pruning Identified Exemplars (PIEs) are systematically more impacted by the introduction of sparsity. Compression disproportionately impacts model performance on the underrepresented long-tail of the data distribution. PIEs over-index on atypical or noisy images that are far more challenging for both humans and algorithms to classify. Our work provides intuition into the role of capacity in deep neural networks and the trade-offs incurred by compression. An understanding of this disparate impact is critical given the widespread deployment of compressed models in the wild.

Key facts. Cited by 171 papers in the map (data through 14 Oct 2026). First posted on arXiv 13 Nov 2019.

Builds on

No reference list yet.

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2021: 27; 2022: 38; 2023: 36; 2024: 28; 2025: 20; 2026: 22.

171 papers.

Sort:

By area:

  1. 2022Matryoshka Representation Learningcited by 282
  2. 2023Model Sparsity Can Simplify Machine Unlearningcited by 214
  3. 2021Deep Learning Through the Lens of Example Difficultycited by 141
  4. 2021Can Subnetwork Structure be the Key to Out-of-Distribution Generalization?cited by 75
  5. 2022The Effects of Regularization and Data Augmentation are Class Dependentcited by 67
  6. 2021A Winning Hand: Compressing Deep Networks Can Improve Out-Of-Distribution Robustnesscited by 62
  7. 2022When does dough become a bagel? Analyzing the remaining mistakes on ImageNetcited by 61
  8. 2022The State of Sparse Training in Deep Reinforcement Learningcited by 61
  9. 2021Randomness In Neural Network Training: Characterizing The Impact of Toolingcited by 59
  10. 2021How Well Do Sparse Imagenet Models Transfer?cited by 50
  11. 2022Mechanistic Mode Connectivitycited by 50
  12. 2021Self-Damaging Contrastive Learningcited by 48
38 more
  1. 2021Meta-Learning Sparse Implicit Neural Representationscited by 48
  2. 2021The Low-Resource Double Bind: An Empirical Study of Pruning for Low-Resource Machine Translationcited by 47
  3. 2022Membership Inference Attacks and Defenses in Neural Network Pruningcited by 45
  4. 2021Efficient Lottery Ticket Finding: Less Data is Morecited by 44
  5. 2022Pruning has a disparate impact on model accuracycited by 44
  6. 2024How Does Quantization Affect Multilingual LLMs?cited by 43
  7. 2021A Primer on Contrastive Pretraining in Language Processing: Methods, Lessons Learned and Perspectivescited by 41
  8. 2022The Combinatorial Brain Surgeon: Pruning Weights That Cancel One Another in Neural Networkscited by 41
  9. 2021BERTnesia: Investigating the capture and forgetting of knowledge in BERTcited by 40
  10. 2022Fortuitous Forgetting in Connectionist Networkscited by 39
  11. 2022Characterizing Datapoints via Second-Split Forgettingcited by 37
  12. 2022Pruning's Effect on Generalization Through the Lens of Training and Regularizationcited by 31
  13. 2024Accuracy is Not All You Needcited by 31
  14. 2022Learning with Recoverable Forgettingcited by 30
  15. 2022FairGRAPE: Fairness-aware GRAdient Pruning mEthod for Face Attribute Classificationcited by 30
  16. 2021The Rich Get Richer: Disparate Impact of Semi-Supervised Learningcited by 28
  17. 2024A Systematic Review on Long-Tailed Learningcited by 28
  18. 2025A Coreset Selection of Coreset Selection Literature: Introduction and Recent Advancescited by 28
  19. 2021Effective Model Sparsification by Scheduled Grow-and-Prune Methodscited by 26
  20. 2022Sparse Double Descent: Where Network Pruning Aggravates Overfittingcited by 26
  21. 2022Metadata Archaeology: Unearthing Data Subsets by Leveraging Training Dynamicscited by 26
  22. 2022Controlled Sparsity via Constrained Optimization or: How I Learned to Stop Tuning Penalties and Love Constraintscited by 25
  23. 2023Interpretability Illusions in the Generalization of Simplified Modelscited by 25
  24. 2024Learned feature representations are biased by complexity, learning order, position, and morecited by 23
  25. 2021Keep the Gradients Flowing: Using Gradient Flow to Study Sparse Network Optimizationcited by 22
  26. 2023Model Compression in Practice: Lessons Learned from Practitioners Creating On-device Machine Learning Experiencescited by 21
  27. 2022Selective Prediction via Training Dynamicscited by 20
  28. 2023On the Effectiveness of Out-of-Distribution Data in Self-Supervised Long-Tail Learningcited by 20
  29. 2023Sparse Model Soups: A Recipe for Improved Pruning via Model Averagingcited by 20
  30. 2025Uncovering Memorization Effect in the Presence of Spurious Correlationscited by 20
  31. 2021Scaling Up Exact Neural Network Compression by ReLU Stabilitycited by 19
  32. 2021A Tale Of Two Long Tailscited by 19
  33. 2022Deconstructing Distributions: A Pointwise Framework of Learningcited by 19
  34. 2023Understanding the Effect of Model Compression on Social Bias in Large Language Modelscited by 19
  35. 2024On the Limitations of Compute Thresholds as a Governance Strategycited by 18
  36. 2022Towards Fair Federated Recommendation Learning: Characterizing the Inter-Dependence of System and Data Heterogeneitycited by 17
  37. 2023Adaptive Sharpness-Aware Pruning for Robust Sparse Networkscited by 17
  38. 2023Late Stopping: Avoiding Confidently Learning from Mislabeled Examplescited by 17

50 most cited of 171 All 171 (JSON)