Mitigating task interference in multi-task learning via gradient-guided sparse modeling

Yayu Zhang

doi:10.1007/s13042-026-03366-0

Mitigating task interference in multi-task learning via gradient-guided sparse modeling

2026-11-05

0

PRE

AI

DOI：10.1007/s13042-026-03366-0

Original

Save

Abstract

En 中文

Task interference remains a fundamental challenge in multi-task learning (MTL), limiting the effectiveness of joint optimization across multiple related tasks. Existing approaches to mitigating task interference mainly fall into two categories. Sparse multi-task learning (SMTL) methods learn task-specific sparse masks by estimating parameter importance independently to partition the shared backbone. However, such method neglects cross-task parameter dependencies and dynamic gradient conflicts, resulting in suboptimal mask allocation. Optimization-based MTL methods alleviate task interference by seeking a compromise among conflicting task gradients. However, such compromise often leads to suboptimal solutions for individual tasks, limiting overall multi-task performance. To overcome the limitations of both methods, we propose GD-SMTL, a gradient-dependence-guided SMTL framework that jointly exploits parameter utility and cross-task gradient interactions for sparse parameter allocation. Specifically, GD-SMTL estimates task-specific parameter utility using robustly normalized Fisher information and captures persistent cross-task gradient interactions from exponential moving average gradient trajectories. An asymmetric dependency-aware allocation strategy then assigns contested parameters according to relative task dependence, enabling more effective parameter allocation and mitigating task interference. Extensive experiments on diverse benchmarks, including multi-task image classification, multi-label face attribute prediction, and dense prediction, demonstrate that GD-SMTL effectively suppresses task interference while promoting positive knowledge transfer.

Keywords:

Sparse multi-task learning

Task interference

Parameter dependence

Gradient conflict

Mitigating task interference in multi-task learning via gradient-guided sparse modeling

Abstract

Journal

Organization

Cited Papers