mlsanityData to 7 Oct 2026

Anatomy of Catastrophic Forgetting: Hidden Representations and Task Semantics

arXiv 2007.07400 ↗First posted 14 Jul 2020cs.LG cs.CV stat.MLcited by 158

Not on the map. Why?

Abstract (arXiv)

A central challenge in developing versatile machine learning systems is catastrophic forgetting: a model trained on tasks in sequence will suffer significant performance drops on earlier tasks. Despite the ubiquity of catastrophic forgetting, there is limited understanding of the underlying process and its causes. In this paper, we address this important knowledge gap, investigating how forgetting affects representations in neural network models. Through representational analysis techniques, we find that deeper layers are disproportionately the source of forgetting. Supporting this, a study of methods to mitigate forgetting illustrates that they act to stabilize deeper layers. These insights enable the development of an analytic argument and empirical picture relating the degree of forgetting to representational similarity between tasks. Consistent with this picture, we observe maximal forgetting occurs for task sequences with intermediate similarity. We perform empirical studies on the standard split CIFAR-10 setup and also introduce a novel CIFAR-100 based task approximating realistic input distribution shift.

Key facts. Cited by 158 papers in the map (data through 7 Oct 2026). First posted on arXiv 14 Jul 2020.

Builds on

No reference list yet.

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2021: 17; 2022: 27; 2023: 31; 2024: 39; 2025: 19; 2026: 25.

158 papers.

Sort:

By area:

  1. 2021A General Language Assistant as a Laboratory for Alignmentcited by 1,023
  2. 2023A Comprehensive Survey of Continual Learning: Theory, Method and Applicationcited by 860
  3. 2022Surgical Fine-Tuning Improves Adaptation to Distribution Shiftscited by 205
  4. 2021An Empirical Investigation of the Role of Pre-training in Lifelong Learningcited by 134
  5. 2021Achieving Forgetting Prevention and Knowledge Transfer in Continual Learningcited by 91
  6. 2021Continual Learning in the Teacher-Student Setup: Impact of Task Similaritycited by 86
  7. 2024How Do Large Language Models Acquire Factual Knowledge During Pretraining?cited by 86
  8. 2022Reliability of CKA as a Similarity Measure in Deep Learningcited by 82
  9. 2022How catastrophic can catastrophic forgetting be in linear regression?cited by 80
  10. 2022Probing Representation Forgetting in Supervised and Unsupervised Continual Learningcited by 74
  11. 2022Teaching Structured Vision&Language Concepts to Vision&Language Modelscited by 71
  12. 2023Theory on Forgetting and Generalization of Continual Learningcited by 69
38 more
  1. 2022Continual Learning with Foundation Models: An Empirical Study of Latent Replaycited by 61
  2. 2022A Closer Look at Rehearsal-Free Continual Learningcited by 54
  3. 2023Achieving a Better Stability-Plasticity Trade-off via Auxiliary Networks in Continual Learningcited by 41
  4. 2022Continual Normalization: Rethinking Batch Normalization for Online Continual Learningcited by 40
  5. 2021Self-Supervised Pre-Training for Transformer-Based Person Re-Identificationcited by 36
  6. 2023The Tunnel Effect: Building Data Representations in Deep Neural Networkscited by 34
  7. 2022The Effect of Task Ordering in Continual Learningcited by 30
  8. 2022Analysis of Catastrophic Forgetting for Random Orthogonal Transformation Tasks in the Overparameterized Regimecited by 28
  9. 2021Kernel Continual Learningcited by 27
  10. 2022How to Reuse and Compose Knowledge for a Lifetime of Tasks: A Survey on Continual Learning and Functional Compositioncited by 27
  11. 2023Degeneration-Tuning: Using Scrambled Grid shield Unwanted Concepts from Stable Diffusioncited by 27
  12. 2024The Joint Effect of Task Similarity and Overparameterization on Catastrophic Forgetting -- An Analytical Modelcited by 26
  13. 2024CLAP4CLIP: Continual Learning with Probabilistic Finetuning for Vision-Language Modelscited by 25
  14. 2024Fine-tuning Reinforcement Learning Models is Secretly a Forgetting Mitigation Problemcited by 23
  15. 2023A Survey on Continual Semantic Segmentation: Theory, Challenge, Method and Applicationcited by 21
  16. 2022Challenging Common Assumptions about Catastrophic Forgettingcited by 20
  17. 2025OPLoRA: Orthogonal Projection LoRA Prevents Catastrophic Forgetting during Parameter-Efficient Fine-Tuningcited by 20
  18. 2021How do Quadratic Regularizers Prevent Catastrophic Forgetting: The Role of Interpolationcited by 19
  19. 2024Disentangling and Mitigating the Impact of Task Similarity for Continual Learningcited by 19
  20. 2022Sparsity and Heterogeneous Dropout for Continual Learning in the Null Space of Neural Activationscited by 18
  21. 2023Does Continual Learning Equally Forget All Parameters?cited by 18
  22. 2024ReSi: A Comprehensive Benchmark for Representational Similarity Measurescited by 18
  23. 2022Maslow's Hammer for Catastrophic Forgetting: Node Re-Use vs Node Activationcited by 16
  24. 2023Knowledge Accumulation in Continually Learned Representations and the Issue of Feature Forgettingcited by 16
  25. 2025Unlocking the Power of Function Vectors for Characterizing and Mitigating Catastrophic Forgetting in Continual Instruction Tuningcited by 16
  26. 2021Lifelong Learning with Sketched Structural Regularizationcited by 15
  27. 2021Learning Curves for Continual Learning in Neural Networks: Self-Knowledge Transfer and Forgettingcited by 14
  28. 2023Layer-wise Auto-Weighting for Non-Stationary Test-Time Adaptationcited by 14
  29. 2024Optimal Protocols for Continual Learning via Statistical Physics and Control Theorycited by 14
  30. 2025Optimal Task Order for Continual Learning of Multiple Taskscited by 14
  31. 2023Overcoming the Stability Gap in Continual Learningcited by 13
  32. 2024What Variables Affect Out-of-Distribution Generalization in Pretrained Models?cited by 13
  33. 2024Order parameters and phase transitions of continual learning in deep neural networkscited by 13
  34. 2022Modelling continual learning in humans with Hebbian context gating and exponentially decaying task signalscited by 12
  35. 2022Optimizing Class Distribution in Memory for Multi-Label Online Continual Learningcited by 12
  36. 2022Continual learning: a feature extraction formalization, an efficient algorithm, and fundamental obstructionscited by 11
  37. 2023Revisiting Single-gated Mixtures of Expertscited by 10
  38. 2024Accelerating Heterogeneous Federated Learning with Closed-form Classifierscited by 10

50 most cited of 158 All 158 (JSON)