arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2411.03055cs.LGcs.AIcs.CV
收藏

ATM:通过交替调优与合并改进模型合并

ATM: Improving Model Merging by Alternating Tuning and Merging

Sapienza University of Rome(萨皮恩扎罗马大学)

Luca Zhou, Daniele Solombrino, Donato Crisostomi, Maria Sofia Bucarelli, Fabrizio Silvestri, Emanuele Rodolà

AI总结:

本文提出ATM方法,通过交替调优与合并改进模型合并,理论证明任务向量等价于多任务梯度,并在数据受限场景及轻量级精炼中验证其有效性。

摘要:

模型合并已成为多任务学习的一种经济高效的近似方法。在各类合并策略中,任务算术因其简单性和有效性而备受关注。在本工作中,我们为任务向量提供了理论动机,指出在单轮全批量梯度下降下,任务向量等价于多任务梯度。这一见解促使我们将模型合并重新解释为一种迭代过程中的单一步骤,该过程在调优与合并之间交替进行(ATM)。我们提出了ATM的两种应用:(1)作为数据共享受限场景(如联邦设置)中多任务学习的替代方案;(2)作为利用小型验证集改进现有模型合并方法的轻量级精炼步骤。跨多种视觉任务的实验证明了ATM的有效性。

英文摘要:

Model merging has emerged as a cost-efficient approximation to multitask learning. Among merging strategies, task arithmetic is notable for its simplicity and effectiveness. In this work, we provide a theoretical motivation for task vectors by highlighting that, under single-epoch full-batch gradient descent, they are equivalent to multitask gradients. This insight leads us to reinterpret model merging as a single step in an iterative procedure that Alternates between Tuning and Merging (ATM). We propose two applications of ATM: (1) as an alternative to multitask learning in scenarios where data sharing is restricted (e.g., federated settings), and (2) as a lightweight refinement step to improve existing model merging methods using a small validation set. Experiments across diverse vision tasks demonstrate the effectiveness of ATM.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑