arXivDaily arXiv每日学术速递 周一至周五更新

大厂专区

Intel(英特尔)

2026-05-01 至 2026-05-01 共收录 2
2310.02277 2026-05-01 cs.LG cs.AI

Junk DNA Hypothesis: Pruning Small Pre-Trained Weights Irreversibly and Monotonically Impairs "Difficult" Downstream Tasks in LLMs

垃圾DNA假说:修剪小的预训练权重不可逆且单调地损害LLM中的“困难”下游任务

Lu Yin, Ajay Jaiswal, Shiwei Liu, Souvik Kundu, Zhangyang Wang

机构 * University of Surrey Eindhoven University of Technology University of Texas at Austin Intel Labs University of Oxford

AI总结 该研究提出垃圾DNA假说,指出LLM预训练权重中存在关键知识,修剪小权重会单调损害困难下游任务性能,且即使允许持续训练也无法弥补损失。

Comments Published at ICML 2024

详情

展开后加载摘要…

URL PDF HTML 收藏
2604.27562 2026-05-01 cs.LG

Online semi-supervised perception: Real-time learning without explicit feedback

在线半监督感知:无需显式反馈的实时学习

Branislav Kveton, Michal Valko, Matthai Phillipose, Ling Huang

机构 * Intel Labs(英特尔实验室) Department of Computer Science(计算机科学系) University of Pittsburgh(匹兹堡大学)

AI总结 本文提出一种无需显式反馈的实时学习算法,结合图上半监督学习和在线学习,通过迭代构建世界图表示并更新,利用离线标注数据和在线未标注数据提升性能,在实时人脸识别中取得优越精度和召回率。

Comments IEEE Computer Vision and Pattern Recognition Workshop on Online Learning for Computer Vision (CVPR 2010 OLCV)

详情

展开后加载摘要…

URL PDF HTML 收藏