New AI Unlearning Method Reduces Multi-Task Interference by 53%

Machine unlearning—the ability to erase specific training data from a trained model—has become a critical tool for privacy and compliance, but existing techniques largely assume single-task models. Modern AI systems often use shared backbones across multiple tasks (e.g., image classification, segmentation, and detection simultaneously), meaning removing data for one task can inadvertently degrade performance on others. Researchers from National Taiwan University (Huang, Fang, Chen, Chen) formalize this problem as “multi-task unlearning” with two variants: full-task unlearning (erase an instance from all tasks) and partial-task unlearning (erase supervision for an instance only on selected tasks). They identify two types of interference—task-level (non-target tasks affected) and instance-level (other data points harmed)—and show that naive unlearning methods fail to handle the coupling caused by shared parameters.

To solve this, the team proposes an interference-aware framework that combines two mechanisms: task-aware gradient projection, which constrains weight updates to task-specific subspaces to avoid cross-task bleeding, and instance-level gradient orthogonalization, which aligns forget and retain signals to minimize conflicts. Tested on two multi-task computer vision benchmarks spanning five tasks, the method consistently outperforms existing baselines. Key result: unintended interference score (UIS) dropped 30.3% for full-task unlearning and 52.9% for partial-task unlearning relative to the strongest baseline. While the paper focuses on vision tasks, the principles are general and could extend to large language models and multi-modal systems. This work provides a principled foundation for deploying unlearning in production multi-task AI, enabling safer data removal without sacrificing overall model utility.