ManiF-SMC: A New Approach to Machine Unlearning via Manifold Forgetting
A recent study introduces ManiF-SMC (Manifold Forgetting with Self Mode Connectivity) aimed at facilitating machine unlearning in support of the right to be forgotten. Current techniques that depend on altering labels or reversing task gradients frequently fall short and may compromise the initial learning goals. ManiF-SMC redefines approximate unlearning by moving each removed sample away from its original learned manifold centroid and towards its closest semantic neighbors within the retained dataset, thus aligning unlearning with the process of retraining. This research can be found on arXiv with the ID 2605.22871.
Key facts
- ManiF-SMC stands for Manifold Forgetting with Self Mode Connectivity.
- The paper addresses machine unlearning for the right to be forgotten.
- Existing unlearning methods using label manipulation or task-gradient reversal have limited effectiveness.
- ManiF-SMC pushes erased samples away from their original manifold centroid.
- The approach aligns unlearning with retraining on remaining data.
- The paper is published on arXiv with ID 2605.22871.
Entities
Institutions
- arXiv