Interference-Aware Multi-Task Unlearning
The paper introduces a framework for multi-task unlearning in machine learning, addressing the challenges of removing specific training data without affecting other tasks. It presents two unlearning settings: full-task and partial-task unlearning, and highlights the interference caused by shared parameters. The proposed method demonstrates significant improvements in unlearning effectiveness while maintaining model generalization across multiple tasks.
- ▪Machine unlearning aims to remove designated training data from a model while preserving performance on remaining data.
- ▪The proposed framework combines task-aware gradient projection with instance-level gradient orthogonalization to reduce interference.
- ▪Experiments show a 30.3% reduction in unlearning interference score for full-task unlearning and 52.9% for partial-task unlearning.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Artificial Intelligence arXiv:2605.19042 (cs) [Submitted on 18 May 2026] Title:Interference-Aware Multi-Task Unlearning Authors:Ying-Hua Huang, Rui Fang, Hsi-Wen Chen, Ming-Syan Chen View a PDF of the paper titled Interference-Aware Multi-Task Unlearning, by Ying-Hua Huang and 3 other authors View PDF HTML (experimental) Abstract:Machine unlearning aims to remove the contribution of designated training data from a trained model while preserving performance on the remaining data. Existing work mainly focuses on single-task settings, whereas modern models often operate in multi-task setups with shared backbones, where removing supervision for one task or instance can unintentionally affect others.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
