Deakin University
Browse

File(s) under permanent embargo

Dimensionality-Driven learning with noisy labels

Version 2 2024-06-06, 10:41
Version 1 2020-07-08, 14:43
conference contribution
posted on 2024-06-06, 10:41 authored by X Ma, Y Wang, ME Houle, S Zhou, SM Erfani, ST Xia, S Wijewickrema, J Bailey
© The Author(s) 2018. 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 represen-tation subspace of training samples. We show that from a dimensionality perspective, DNNs ex-hibit quite distinctive learning styles when trained with clean labels versus when trained with a pro-portion 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.

History

Volume

8

Pagination

5332-5341

Location

Stockholm, Sweden

Start date

2018-07-10

End date

2018-07-15

ISBN-13

9781510867963

Language

eng

Publication classification

E1.1 Full written paper - refereed

Title of proceedings

ICML 2018 : Proceedings of the 35th International Conference on Machine Learning

Event

Machine Learning. Conference (2018 : 35th : Stockholm, Sweden)

Publisher

ICML

Place of publication

[Stockholm, Sweden]

Usage metrics

    Research Publications

    Exports

    RefWorks
    BibTeX
    Ref. manager
    Endnote
    DataCite
    NLM
    DC