Papers13:43Chinese content研究:大步长梯度下降在对抗训练中的隐式偏差与损失锐度行为搞对抗训练的人可以看看这篇理论分析,它证明了用大步长训练也照样成立,不用非得用慢吞吞的小步长。#adversarial training#gradient descent#model robustnessaarXiv cs.LG@Yi Feng, Andrea PaudiceSourceRead laterReadWorth followingUsefulFollow adversarial training
Papers10:49Chinese contentOOD退化预测研究:源训练动力学可提前预警这篇论文发现源训练动力学能提前预警未来OOD失效,对模型鲁棒性研究有实用价值。#OOD#training dynamics#shortcut learning#model robustnessaarXiv cs.LG@ Sasha, MoninSourceRead laterReadWorth followingUsefulFollow OOD
PapersBuilders10:21Chinese content预测价值表示对LLM对齐泛化的影响这篇论文研究了LLM价值对齐的泛化预测,用激活状态表示比文本描述效果好9倍,还能评估多价值目标的鲁棒性。#LLM#alignment#generalization#model robustnessaarXiv cs.AI@Andy Liu et al. (6)SourceRead laterReadWorth followingUsefulFollow LLM
Papers11:10Chinese contentAutoBackSwap:自动背景交换提升分类器对虚假背景的鲁棒性这篇论文提出了AutoBackSwap,只用几百张图片的标注就能让模型不依赖背景,效果比之前的方法都好,做图像分类的值得一看。#AutoBackSwap#image classification#data augmentation#model robustnessaarXiv cs.LG@Cesar Roder, Kajetan SchweighoferSourceRead laterReadWorth followingUsefulFollow AutoBackSwap