Interference-Aware Multi-Task Unlearning Framework
A new arXiv paper (2605.19042) introduces multi-task machine unlearning, addressing interference in models with shared backbones. The authors propose two settings: full-task unlearning (removing a target instance from all tasks) and partial-task unlearning (removing supervision from selected tasks). They identify task-level and instance-level interference due to shared parameters, and present an interference-aware framework combining task-aware gradient projection and instance-level gradient orthogonalization.
Key facts
- Paper title: Interference-Aware Multi-Task Unlearning
- arXiv ID: 2605.19042
- Announce type: new
- Focuses on multi-task unlearning
- Two settings: full-task and partial-task unlearning
- Identifies task-level and instance-level interference
- Proposes interference-aware framework with gradient projection and orthogonalization
- Published on arXiv
Entities
Institutions
- arXiv