Towards Task-Conflicts Momentum-Calibrated Approach for Multi-task Learning

Link
Abstract

Multi-task learning (MTL) has succeeded in various industrial applications by utilizing common knowledge among joint training tasks to enhance the generalization of MTL models, resulting in improved performance across all training tasks simultaneously. Unfortunately, training all tasks simultaneously often causes performance degradation compared to single-task models since different tasks might conflict with each other. Despite existing MTL methods that aim to mitigate task conflicts by manipulating task gradients at each iteration, they ignore the potential influence of noisy data from different batches on task gradients. Consequently, the current iteration's task gradient may not accurately reflect the task itself, leading to inadequate alleviation of the dilemma of task conflicts. Moreover, existing works seldom explore the potential source of task conflicts and merely pose an assumption. In this paper, we conduct an in-depth empirical investigation into the potential sources of performance degradation of MTL and find that task gradient conflict is one of the primary reasons for the performance degradation of tasks. Then, to address the task conflicts problem, we propose a novel gradient manipulation approach, namely MoCoGrad, which manipulates task gradients by leveraging the momentum information of the task to calibrate the gradients of conflicting tasks. In addition, we derive theoretical guarantees for the convergence of our proposed MoCoGrad and theoretically analyze the convergence rate of MoCoGrad. Finally, to evaluate the effectiveness of MoCoGrad, extensive experiments are conducted on six real-world datasets from different domains. Our approach yields the best performance across all tasks in all six MTL benchmarks, demonstrating the effectiveness and superiority of our method.

Synth

Problem:: 기존 연구들은 Gradient Conflict를 MTL 성능 저하의 원인으로 지목하면서도 분석하지 않음/기존 방법들이 현재 Iteration의 Gradient만 고려하여 Noisy Data 영향에 취약함

Solution:: Task Conflict와 Gradient Conflict의 관게 규명/Task의 Momentum 정보를 활용하여 Conflicting Gradient를 보정(calibrate)하는 MoCoGrad 제안

Novelty:: 작업 간 충돌 원인에 대한 체계적 분석 및 정량화

Note:: TCI와 GCD를 GazeTargetDetection의 두 Task간의 충돌 정도를 나타내는데 이용해 볼 수 있을듯

Summary

Motivation

HPS와 MMoE는 MTL을 위한 아키텍쳐

Method

Task Conflicts Analysis

file-20250327164130301.png

Task Conflict와 Gradient Conflict가 강한 상관관계를 보임 → 우리가 처음으로 이거 보였음! 다른 애들은 그냥 대강 보이더라

Momentum-calibrated Conflicting Gradients (MoCoGrad)

file-20250327164346891.png

기존 방식은 한 배치 기준으로 계산 → 배치 별로 보정되는 방향이 다른데, 너무 노이즈가 심함

Method 검증