Skip to main content

Addressing the labeled data scarcity problem in deep learning

Resource type
Thesis type
(Thesis) Ph.D.
Date created
2023-08-28
Authors/Contributors
Abstract
A significant challenge in machine learning and, more specifically, deep learning is the necessity for extensive labeled data. Given the high cost, time, and complexity of procuring such data, especially in computer vision applications, we focus on improving strategies for scenarios with limited data. We propose novel methodologies in three primary areas: Active Learning, Data Augmentation, and Domain Adaptation. Firstly, in the domain of Active Learning, we propose an algorithm that caters to cases with structured outputs, efficiently expanding the labeled dataset from a large pool of unlabeled data. By exploiting the inherent interdependencies in each sample, we can identify the most informative components for further labeling. Secondly, under Data Augmentation, we propose a unique algorithm for synthesizing realistic human action videos to enhance the training dataset for human action recognition tasks. Lastly, in the area of Domain Adaptation, we address the challenges of varying data distributions between training and deployment stages in object detection. A novel algorithm is proposed to enable object detectors adapt effectively using available unlabeled data from the target domain.
Document
Extent
100 pages.
Identifier
etd22712
Copyright statement
Copyright is held by the author(s).
Permissions
This thesis may be printed or downloaded for non-commercial research and scholarly purposes.
Supervisor or Senior Supervisor
Thesis advisor: Mori, Greg
Language
English
Member of collection
Download file Size
etd22712.pdf 10.11 MB

Views & downloads - as of June 2023

Views: 44
Downloads: 8