MTLoRA: Low-Rank Adaptation Approach for Efficient Multi-Task Learning

Link
Abstract

Adapting models pre-trained on large-scale datasets to a variety of downstream tasks is a common strategy in deep learning. Consequently, parameter-efficient fine-tuning methods have emerged as a promising way to adapt pretrained models to different tasks while training only a minimal number of parameters. While most of these methods are designed for single-task adaptation, parameter-efficient training in Multi-Task Learning (MTL) architectures is still unexplored. In this paper, we introduce MTLoRA, a novel framework for parameter-efficient training of MTL models. MTLoRA employs Task-Agnostic and Task-Specific LowRank Adaptation modules, which effectively disentangle the parameter space in MTL fine-tuning, thereby enabling the model to adeptly handle both task specialization and interaction within MTL contexts. We applied MTLoRA to hierarchical-transformer-based MTL architectures, adapting them to multiple downstream dense prediction tasks. Our extensive experiments on the PASCAL dataset show that MTLoRA achieves higher accuracy on downstream tasks compared to fully fine-tuning the MTL model while reducing the number of trainable parameters by 3.6×. Furthermore, MTLoRA establishes a Pareto-optimal trade-off between the number of trainable parameters and the accuracy of the downstream tasks, outperforming current stateof-the-art parameter-efficient training methods in both accuracy and efficiency. Our code is publicly available.

Synth

Problem:: LoRA는 효과적이지만 MTL에서 효과가 떨어짐 / 각 Task별로 새로운 LoRA를 도입하는 것은 연산량을 증가시킴

Solution:: MTL 시나리오에 맞는 Task Specific, Task Agnostic LoRA 모듈 제안 / 각 Encoder Stage 마지막에만 Task 별 LoRA를 도입해 연산량 증가 최소화

Novelty:: MTL 시나리오에서 LoRA를 처음으로 사용

Note:: MultiLoRA 논문이 LoRA가 왜 FT와 다르며, 이 차이 때문에 성능 저하가 발생하고 이를 이를 해결하기 위해 LoRA의 수평적 병렬화를 제안했다면 이 논문은 기존 MTL 시나리오에서 Task별 모듈을 여러개 두는 것에 LoRA를 적용시키고, 이로인한 연산량 증가를 TA, TS로 구분지어서 줄였음. 전형적인 SOTA 찍었으니 왜 이 메소드가 좋은지는 굳이 설명하지 않음 느낌

Summary

Motivation

Method

file-20250407125913597.png

Method 검증

실험 환경

MTLoRA vs. Full Fine-tuning MTL

MTLoRA의 Rank(r) 영향

MTLoRA+의 효율성

Non-Attention Module 영향 분석

Low-Rank Decomposition Module 위치에 따른 영향

대규모 Backbone에서의 확장성