Machine Unlearning
Techniques to remove the influence of specific training data from an ML model without retraining the entire model.
Machine unlearning removes the influence of specific data from trained models – essential for GDPR compliance (right to erasure).
Explanation
Exact unlearning retrains from scratch (expensive). Approximate unlearning uses gradient-based methods, SISA training, or influence functions to efficiently remove individual data point influence.
Marketing Relevance
GDPR Article 17 (right to erasure) requires not just data deletion but also removal from trained models – machine unlearning makes this practical.
Example
A user requests data deletion under GDPR. The company deletes their data AND removes their influence from the recommendation model via SISA training.
Common Pitfalls
Verifying unlearning is difficult. Approximate unlearning doesn't offer perfect guarantees. For LLMs, unlearning is particularly challenging.
Origin & History
Cao & Yang (2015) formalized machine unlearning. SISA Training (Bourtoule et al., 2021) made it practical. Google's Machine Unlearning Challenge (2023) advanced research. LLM unlearning is an active research area.
Comparisons & Differences
Machine Unlearning vs. Differential Privacy
DP prevents individual data from being identifiable from the start; unlearning removes data retroactively from the model.
Machine Unlearning vs. Data Deletion
Data deletion removes raw data; unlearning additionally removes the learned influence of that data from the model.