Dimensionality-Driven Learning with Noisy Labels-猫眼课题宝

DOI：

发表时间：

2018-06

期刊：

ArXiv

影响因子：

通讯作者：

中科院分区：

其他

文献类型：

作者：

Xingjun Ma;Yisen Wang;Michael E. Houle;Shuo Zhou;S. Erfani;Shutao Xia;S. Wijewickrema;J. Bailey

具有显著比例的噪声（不正确）类标签的数据集对训练准确的深度神经网络（dnn）提出了挑战。我们通过研究训练样本的深度表示子空间的维度，为理解DNN泛化提供了一个新的视角。我们表明，从维度的角度来看，dnn在使用干净标签训练时与使用一定比例的噪声标签训练时表现出相当独特的学习风格。基于这一发现，我们开发了一种新的维度驱动学习策略，该策略在训练过程中监控子空间的维度并相应地调整损失函数。我们的经验表明，我们的方法对显著比例的噪声标签具有高度的容忍度，并且可以有效地学习捕获数据分布的低维局部子空间。

Datasets with significant proportions of noisy (incorrect) class labels present challenges for training accurate Deep Neural Networks (DNNs). We propose a new perspective for understanding DNN generalization for such datasets, by investigating the dimensionality of the deep representation subspace of training samples. We show that from a dimensionality perspective, DNNs exhibit quite distinctive learning styles when trained with clean labels versus when trained with a proportion of noisy labels. Based on this finding, we develop a new dimensionality-driven learning strategy, which monitors the dimensionality of subspaces during training and adapts the loss function accordingly. We empirically demonstrate that our approach is highly tolerant to significant proportions of noisy labels, and can effectively learn low-dimensional local subspaces that capture the data distribution.

Dimensionality-Driven Learning with Noisy Labels

文献摘要

被引文献

相似文献