语言模型持续学习中的虚假遗忘
Spurious Forgetting in Continual Learning of Language Models
AI总结：
本研究提出“虚假遗忘”概念，指出持续学习中性能下降源于任务对齐破坏而非知识丢失，并通过冻结底层策略在四个场景中显著提升性能。
摘要：
近期大语言模型（LLMs）的进展揭示了持续学习中的一个令人困惑的现象：尽管经过大量训练，模型仍会出现显著的性能下降，这引发了对任务对齐和底层知识保留的疑问。本研究首先探讨了“虚假遗忘”的概念，提出这种性能下降往往反映的是任务对齐的衰退，而非真正的知识丢失。通过使用合成数据集进行受控实验，我们研究了新任务初始训练阶段模型性能的动态变化，发现早期的优化步骤可能破坏先前建立的任务对齐。我们的理论分析将这些变化与模型权重的正交更新联系起来，为理解这一行为提供了稳健的框架。最终，我们引入了一种冻结策略，固定模型底层，在四个持续学习场景中带来了显著的改进。我们的发现强调了任务对齐与知识保留之间的关键区别，为持续学习中更有效的策略铺平了道路。
英文摘要：
Recent advancements in large language models (LLMs) reveal a perplexing phenomenon in continual learning: despite extensive training, models experience significant performance declines, raising questions about task alignment and underlying knowledge retention. This study first explores the concept of "spurious forgetting", proposing that such performance drops often reflect a decline in task alignment rather than true knowledge loss. Through controlled experiments with a synthesized dataset, we investigate the dynamics of model performance during the initial training phases of new tasks, discovering that early optimization steps can disrupt previously established task alignments. Our theoretical analysis connects these shifts to orthogonal updates in model weights, providing a robust framework for understanding this behavior. Ultimately, we introduce a Freezing strategy that fix the bottom layers of the model, leading to substantial improvements in four continual learning scenarios. Our findings underscore the critical distinction between task alignment and knowledge retention, paving the way for more effective strategies in continual learning.
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。
