arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2408.04140cs.CLcs.AI
收藏

UNLEARN:大语言模型中知识的高效移除

UNLEARN Efficient Removal of Knowledge in Large Language Models

Georgia Institute of Technology(佐治亚理工学院)

Tyler Lizzo, Larry Heck

AI总结:

本文提出 UNLEARN 子空间方法,可在不重训大模型的情况下精准遗忘 96% 目标知识且对其他知识影响小于 2.5%,并提出对偶方法 LEARN 实现可媲美 LoRA 的定向知识添加。

摘要:

鉴于大语言模型(LLMs)的普及以及从头训练这些模型的高昂成本,在不重新训练模型的情况下动态遗忘特定知识(例如私有或专有知识)已成为一项重要能力。本文提出了一种名为 UNLEARN 的新方法来实现这一目标。该方法基于子空间方法,识别并专门针对性地移除知识,同时不会对大语言模型中的其他知识产生不利影响。结果表明,96% 的目标知识可以被遗忘,同时在其他知识上的性能保持在原始模型的 2.5% 以内,显著优于先前最先进方法的判别能力。此外,还提出了一种名为 LEARN 的对偶方法,用于有针对性地添加知识。结果表明,LEARN 能够达到与低秩适应(LoRA)微调相当的准确率,且不会对相似任务产生不利影响。

英文摘要:

Given the prevalence of large language models (LLMs) and the prohibitive cost of training these models from scratch, dynamically forgetting specific knowledge e.g., private or proprietary, without retraining the model has become an important capability. This paper proposes a novel method to achieve this objective called UNLEARN. The approach builds upon subspace methods to identify and specifically target the removal of knowledge without adversely affecting other knowledge in the LLM. Results demonstrate 96% of targeted knowledge can be forgotten while maintaining performance on other knowledge within 2.5% of the original model, significantly outperforming the discriminatory abilities of the previous state-of-the-art. A dual method called LEARN is also proposed for targeted knowledge addition. Results show LEARN can match the fine-tuning accuracy of Low-Rank Adaptation (LoRA) without adversely affecting similar tasks.

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。

↑