Interference-Aware Multi-Task Unlearning

2026-05-18

AI Summary

This work addresses the interference problem in multi-task learning caused by removing supervision signals for specific tasks or instances, which—due to parameter sharing—adversely affects other tasks or samples and degrades model performance. The paper presents the first systematic formulation of task-level and instance-level interference in multi-task forgetting and introduces an interference-aware decoupled optimization framework. This framework constrains parameter updates within task-specific subspaces via task-aware gradient projection and mitigates conflict between forgetting and retention signals through instance-level gradient orthogonalization. Evaluated on two multi-task vision benchmarks, the method substantially outperforms existing approaches, reducing the UIS metric by 30.3% under full-task forgetting and by 52.9% under partial-task forgetting, while maintaining strong generalization capability.

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. We introduce multi-task unlearning with two settings: full-task unlearning, which removes a target instance from all tasks, and partial-task unlearning, which removes supervision only from selected tasks. We show that shared parameters couple the forget and retain sets, causing task-level interference on non-target tasks and instance-level interference on other instances. To address this issue, we propose an interference-aware framework that combines task-aware gradient projection, which constrains updates within task-specific subspaces, with instance-level gradient orthogonalization, which reduces conflicts between forget and retain signals. Experiments on two multi-task computer vision benchmarks across five tasks show that our method achieves effective unlearning while maintaining strong generalization, reducing UIS compared with the strongest baseline by 30.3% in full-task unlearning and 52.9% in partial-task unlearning.

Problem

• machine unlearning
• multi-task learning
• task interference
• shared parameters
• data removal

Innovation

• multi-task unlearning
• interference-aware
• gradient projection
• gradient orthogonalization
• machine unlearning