Training with Confidence: 用自动化主动检查捕捉深度学习训练中的隐式错误
机器学习
2025-06-19 v1 人工智能
摘要
深度学习模型的训练是一个复杂的过程，容易引入难以检测和诊断的隐式错误。本文提出了TRAINCHECK框架，采用主动检查方法以应对隐式训练错误。TRAINCHECK自动推断针对深度学习训练的不变量，并利用这些不变量在训练过程中主动检测隐式错误，同时提供调试支持。为评估TRAINCHECK，我们复现了20个具有多样化根源的真实隐式训练错误。TRAINCHECK在单个训练迭代中成功检测了18个错误。它还发现了6个在流行训练库中导致隐式错误的未知bug。
关键词
引用
@article{arxiv.2506.14813,
title = {Training with Confidence: Catching Silent Errors in Deep Learning Training with Automated Proactive Checks},
author = {Yuxuan Jiang and Ziming Zhou and Boyu Xu and Beijie Liu and Runhui Xu and Peng Huang},
journal= {arXiv preprint arXiv:2506.14813},
year = {2025}
}
备注
19 pages, to appear in 19th USENIX Symposium on Operating Systems Design and Implementation (OSDI '25)