Timezone: »

Dimensionality-Driven Learning with Noisy Labels
Xingjun Ma · Yisen Wang · Michael E. Houle · Shuo Zhou · Sarah Erfani · Shutao Xia · Sudanthi Wijewickrema · James Bailey

Thu Jul 12 05:30 AM -- 05:50 AM (PDT) @ A6

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.

Author Information

Daniel Ma (The University of Melbourne)

I am a 3rd-year Ph.D. student at The University of Melbourne, working on machine learning, adversarial deep learning and computer vision.

Yisen Wang (Tsinghua University)
Michael E. Houle (National Institute of Informatics)
Shuo Zhou (The University of Melbourne)

I'm a third year Ph.D student from the University of Melbourne, Australia. I'm studying in the Department of Computing and Information Systems, under the supervision of Prof. James Bailey, Dr. Sarah M. Erfani and Dr. Vinh Xuan Nauyen (previous co-supervisor). My research generally focuses on data mining and machine learning techniques. Prior to starting my Ph.D studying, I've completed a Master's degree in Distributed Computing at the University of Melbourne and received my Bachelor degree in Software Engineering from Huazhong University of Science and Technology, China.

Sarah Erfani (University of Melbourne)
Shutao Xia (Tsinghua University)
Sudanthi Wijewickrema (University of Melbourne)
James Bailey (The University of Melbourne)

Related Events (a corresponding poster, oral, or spotlight)

More from the Same Authors