DataTerminology.comNEW
A neural network’s tendency to abruptly forget previously learned tasks when trained sequentially on new tasks.
DESCRIPTION
Catastrophic forgetting occurs when models trained sequentially overwrite parameters crucial for earlier tasks, degrading prior performance. It is pervasive in continual learning settings without access to all historical data. Mitigation strategies include regularization (e.g., EWC), rehearsal via replay buffers, architectural isolation or expansion, and knowledge distillation. Proper evaluation distinguishes task-, domain-, and class-incremental scenarios. Addressing forgetting is essential for safe long-lived AI systems.
ALTERNATIVE DEFINITIONS
Wikipedia
Catastrophic interference, also known as catastrophic forgetting, is the tendency of an artificial neural network to abruptly forget previous information upon learning new information.
Defines the phenomenon and equates interference with forgetting.
arXiv
When artificial neural networks are trained sequentially on multiple tasks, they tend to forget previously learned tasks.
Highlights the sequential training setting that induces forgetting.
RELATED TERMS
Elastic Weight Consolidation Experience Replay Regularization Stability–Plasticity Dilemma Knowledge Distillation Task Incremental Learning Domain Incremental Learning Replay Buffer
COMPETING TERMS
KNOWLEDGE AREAS
AI Data Science MLOps Testing & Validation
TARGET AUDIENCE
AI Engineer Data Scientist MLOps Engineer
EXTERNAL SOURCES
- Wikipedia - Catastrophic interference
- arXiv - EWC paper mitigating forgetting
- arXiv - Empirical study of forgetting
- ContinualAI - Community resources on continual learning
TOP QUESTIONS
Why does Catastrophic Forgetting happen in neural networks?
Shared parameters are updated to fit new tasks, overwriting representations needed for earlier tasks. Without constraints or rehearsal, the model drifts away from prior optima.
Category: Explanation
How can Catastrophic Forgetting be mitigated?
Use regularization (EWC, SI), replay buffers, pseudo-rehearsal, modular architectures, or knowledge distillation. The best approach depends on privacy, data availability, and compute.
Category: Best Practices
What is Elastic Weight Consolidation in this context?
EWC penalizes changes to parameters deemed important to previous tasks, preserving performance while learning new tasks. It estimates importance using Fisher information.
Category: Techniques
How is Catastrophic Forgetting measured?
Common metrics include average accuracy across tasks and backward transfer. Evaluations differ for task-, domain-, and class-incremental setups.
Category: Measurement
Is Catastrophic Forgetting a problem for LLM fine-tuning?
Yes. Sequential fine-tuning can degrade earlier skills; mitigations include LoRA adapters per task, replay, or multi-task training.
Category: Relevancy
What are common mistakes when studying Catastrophic Forgetting?
Using unrealistic task splits, data leakage, or inconsistent evaluation protocols can misrepresent forgetting severity or mitigation effectiveness.
Category: Mistakes
What is the stability–plasticity dilemma in relation to Catastrophic Forgetting?
It describes the trade-off between retaining past knowledge (stability) and learning new information (plasticity). Methods aim to balance both.
Category: Concept
Can rehearsal always be used to avoid Catastrophic Forgetting?
No. Privacy, IP constraints, or storage limits may preclude storing past data. Generative replay or synthetic data may substitute.
Category: Risks
Where should I start learning about Catastrophic Forgetting?
Read the EWC paper, Goodfellow et al.’s empirical study, and recent continual learning surveys. Then replicate replay and regularization baselines.
Category: Getting Started