BriefGPT.xyz
Nov, 2020
针对噪声标签的神经网络鲁棒训练核心集
Coresets for Robust Training of Neural Networks against Noisy Labels
HTML
PDF
Baharan Mirzasoleiman, Kaidi Cao, Jure Leskovec
TL;DR
提出了一种基于coresets和梯度下降的方法来处理具有噪声标签的深度神经网络的鲁棒性训练问题,并证明该方法不会过拟合噪声标签,实验证明该方法取得了诸如在CIFAR-10上以80%噪声标签训练后,准确度提高了6%,在mini Webvision上准确度提高了7%的显着优异表现。
Abstract
Modern
neural networks
have the capacity to overfit
noisy labels
frequently found in real-world datasets. Although great progress has been made, existing techniques are limited in providing theoretical guarantees
→