How can a learnt ML model unlearn something: Framework for "Machine Unlearning"
summary
Survey of methods and limits for removing training influence without full retrain claims a working machine unlearning framework and demos verification by showing weight shifts and attacks. Emphasis is on three concrete criteria: completeness, timeliness, verifiability. Verification examples include a feature-injection set test and membership-inference checks; evaluation uses layer-wise weight distance and unlearn time. Notes most prior work targets linear models, recent papers extend to deep nets since 2020, and retrain cost scales up steeply so unlearning is a cost/accuracy tradeoff.
AI-written summary · about
chapters (11)
- 0:00Machine unlearning intro
- 1:09Why unlearning is necessary
- 3:59Privacy, security, usability drivers
- 5:47History of unlearning frameworks
- 7:54Challenges in unlearning systems
- 10:00General unlearning framework
- 11:25Designing unlearning models
- 13:56Removal request types
- 16:01Verification methods
- 20:19Performance metrics
- 22:48Summary & future directions
related talks
15mLearning and Unlearning Your Data in Federated Settings
18mRectifying Privacy and Efficacy Measurements in Machine Unlearning...
13mTowards Lifecycle Unlearning Commitment Management: Measuring Sample-level...
39mSecure Machine Learning: Overcoming Algorithmic Bias, Adversarial ML in the Wild West of AI
32mThe Importance of Forgetting for Both Humans & Machines
26mHow Machine Learning Can Help
28mWhy we do ML model retraining wrong, and how to do better
32mThe Miseducation of This Machine