← AI Terminology
Machine Unlearning
Machine unlearning aims to remove the influence of specific training data from a deployed model (for privacy, copyright, or user deletion requests) without full retraining from scratch.
An emerging systems and legal compliance problem.
An emerging systems and legal compliance problem.
Why It Matters in AI
Regulations and user rights demand deletion; retraining foundation models is often impossible. Approximate unlearning, data influence methods, and process controls are active research with real enterprise demand.
Key Points
| Aspect | Description |
|---|---|
| Eval | Membership tests; downstream forget quality |
| Goal | Model behaves as if data never seen |
| Hard | Exact unlearning expensive at LLM scale |
| Methods | Retrain subsets, influence functions, gradient ascent, SISA |
| Related | Differential privacy, GDPR erasure, model editing |
| Practice | Often process+legal plus technical approximations |
Simple Analogy
Trying to make a student truly forget one textbook chapter without repeating all of school — approximate at best without a full redo.
Common Usage Examples
- SISA sharded training for easier forgets
- Empirical unlearning on fine-tunes
- Membership inference to audit forgetting
- Policy workflows for deletion requests
Summary
In short: Machine unlearning seeks to erase specific training data’s influence — a hard privacy and compliance frontier for large models.