Machine Unlearningedit
Research topic on removing or correcting data influence from trained models.
Page details
Machine Unlearning studies how to remove, reduce, or correct the effect of selected training data after a model has already been trained. It addresses both privacy and data-centric systems questions: an unlearning method must say what it removes, how faithfully it approximates retraining, and how much computation or latency is saved.1
Introductionedit
The topic covers post-training data operations: certified deletion, exact removal in tree ensembles, and continuous reweighting for fairness or robustness correction. The shared question is whether a trained system can be revised after deployment without simply retraining from scratch each time the data record changes.
Research contextedit
Qiao's research on post-training data operations includes certified deletion, weighted correction, and tree-ensemble updates. It is closely connected to Data Centric ML because the central object is not a new model architecture, but a data operation that changes model behavior. It also connects to Trustworthy AI, since deletion requests, fairness corrections, and robustness interventions are forms of governance over a trained system.
Publicationsedit
| Paper | Venue/status |
|---|---|
| Hessian-Free Online Certified Unlearning | ICLR 2025, 24-28 April 2025, Singapore. |
| DynFrs: An Efficient Framework for Machine Unlearning in Random Forest | ICLR 2025, 24-28 April 2025, Singapore. |
| Beyond Binary Erasure: Soft-Weighted Unlearning for Fairness and Robustness | AAAI 2026, 20-27 January 2026, Singapore. |
Connection to Qiao's workedit
Qiao's unlearning papers cover complementary settings. Hessian-Free Online Certified Unlearning targets certified updates for convex objectives without explicit Hessian inversion. Beyond Binary Erasure generalizes deletion from binary remove-or-keep actions to continuous weights for fairness and robustness. DynFrs: An Efficient Framework for Machine Unlearning in Random Forest studies exact and efficient update mechanisms for random forests. Together they define an arc from mathematical certification to practical low-latency model maintenance.
See alsoedit
Footnotesedit
-
A widely cited formulation is Bourtoule et al., "Machine Unlearning", IEEE Symposium on Security and Privacy 2021, which introduced SISA-style sharded, isolated, sliced, and aggregated training as a practical route to deletion. ↩