Honors College Thesis
 

Volume minimized generalized cross entropy for learning with label noise

Public Deposited

Downloadable Content

Download PDF
https://ir.library.oregonstate.edu/concern/honors_college_theses/sb397h434

Descriptions

Attribute NameValues
Creator
Abstract
  • The types and rates of label noise in real-world data sets present a challenge to machine learning projects. In this thesis, we propose a novel approach to address this issue. Our method combines a noise modelling technique for correcting label noise across the entire data set with a robust loss function to reduce sensitivity to outlier label noise. We suggest using the VolMinNet framework for instance-independent label noise correction and the Generalized Cross Entropy robust loss function for instance-dependent outliers. Additionally, we propose employing volume maximization regularization to improve our label noise model estimations and ensure identifiability. We evaluate our method on benchmark data sets with varying noise levels, including both instance-dependent and instance-independent noise. Experimental results demonstrate the superiority of our approach over existing methods. Our method consistently achieves higher accuracy and robustness in high noise environments while remaining comparable in lower noise environments, showcasing its effectiveness in handling challenging real-world scenarios.
  • Key Words: Machine learning, Deep learning, Label noise, Label noise learning
Resource Type
Date Issued
Degree Level
Degree Name
Degree Field
Degree Grantor
Commencement Year
Advisor
Committee Member
Non-Academic Affiliation
Rights Statement
Publisher
Peer Reviewed
Language
Embargo reason
  • Pending Publication
Embargo date range
  • 2023-06-06 to 2024-01-11

Relationships

Parents:

This work has no parents.

In Collection:

Items

Thumbnail Title Date Uploaded Visibility Actions
file details WolnickDaniel2023.pdf 2023-06-06 Public Download