arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2402.18865cs.LGcs.AIcs.CL
收藏

分析并减少参数高效微调中的灾难性遗忘

Analyzing and Reducing Catastrophic Forgetting in Parameter Efficient Tuning

College of Information Sciences and Technology, Pennsylvania State University(宾夕法尼亚州立大学信息科学与技术学院);School of Computer Science and Technology, Singapore Management University(新加坡管理大学计算机科学与技术学院);School of Computer Science, Hefei Univeristy of Technology(合肥工业大学计算机学院)

Weijieying Ren, Xinlong Li, Lei Wang, Tianxiang Zhao, Wei Qin

AI总结:

本文从模式连通性角度分析LLM持续微调中的灾难性遗忘,并提出基于LoRA参数插值的双记忆回放方法I-LoRA,在八个CL基准上最高提升11%。

摘要:

现有研究表明,大型语言模型(LLM)在语言理解和生成方面表现出卓越性能。然而,当LLM在复杂多样的领域特定下游任务上持续微调时,其在历史任务上的推理性能会急剧下降,这被称为灾难性遗忘问题。学习可塑性与记忆稳定性之间需要保持权衡。许多现有工作已经探索了记忆回放、正则化和参数隔离等策略,但对于持续LLM微调场景中各种相邻极小值之间的几何联系仍知之甚少。在本工作中,我们通过模式连通性的视角研究不同极小值之间的几何联系,模式连通性意味着不同极小值可以通过一条低损失谷连接。通过大量实验,我们揭示了LLM持续学习场景中的模式连通性现象,并发现它能够在可塑性与稳定性之间取得平衡。基于这些发现,我们提出了一种简单而有效的方法,称为基于插值的LoRA(I-LoRA),该方法基于LoRA参数插值构建双记忆经验回放框架。在八个领域特定CL基准上的大量实验和分析表明,I-LoRA相较此前最先进方法持续展现出显著提升,性能增益最高达11%,为未来关于大型语言模型持续学习问题的研究提供了强有力的基线和洞见。我们的代码可在https://github.com/which47/LLMCL获取。

英文摘要:

Existing research has shown that large language models (LLMs) exhibit remarkable performance in language understanding and generation. However, when LLMs are continuously fine-tuned on complex and diverse domain-specific downstream tasks, the inference performance on historical tasks decreases dramatically, which is known as a catastrophic forgetting problem. A trade-off needs to be kept between learning plasticity and memory stability. Plenty of existing works have explored strategies like memory replay, regularization and parameter isolation, but little is known about the geometric connection of various adjacent minima in the continual LLMs fine-tuning scenarios. In this work, we investigate the geometric connections of different minima through the lens of mode connectivity, which means different minima can be connected by a low-loss valley. Through extensive experiments, we uncover the mode connectivity phenomenon in the LLMs continual learning scenario and find that it can strike a balance between plasticity and stability. Building upon these findings, we propose a simple yet effective method called Interpolation-based LoRA (I-LoRA), which constructs a dual-memory experience replay framework based on LoRA parameter interpolations. Extensive experiments and analysis on eight domain-specific CL benchmarks demonstrate that I-LoRA consistently show significant improvement over the previous state-of-the-art approaches with up to $11\%$ performance gains, providing a strong baseline and insights for future research on the large language model continual learning problem. Our code is available at \url{https://github.com/which47/LLMCL}.

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。

↑