跳到正文
原文
arXiv · Statistics Machine Learning· Ciar\'an M. Gilligan-Lee, Joseph Egan, Yuchen Zhu, Michael O'Riordan·· 2 小时前AI 评分49

GRASP:微调期间无监督识别并移除虚假相关

Unsupervised Identification and Removal of Spurious Correlations During Fine-Tuning

AI 导读

研究提出 GRASP(GRadient projection of Associated Spurious Patterns),可在无监督下从朴素 LoRA 微调权重中识别虚假相关对应的潜在因子,并阻止模型新增对其依赖,同时保留预训练内容。

来源:arXiv · Statistics Machine Learning · arxiv.org