Back to search

University of South Florida

Exploring the Use of Enhanced SWAD Towards Building Learned Models that Generalize Better to Unseen Sources

Abstract

dc:description

Deep learning models, typically, take significant time to train. Classifier ensembles are areliable way to increase classifier accuracy and perhaps generalizability to unseen sources of data. These classifiers can be combined with a simple voting scheme. The problem is that having multiple models can very heavily increase training time. Snapshot ensembles have been shown to provide a boost in performance by creating an ensemble of classifiers with different weights during the training of a single deep learned model. This can somewhat solve the problem of the increased training time as you do not have to train separate models. As Machine Learning becomes a more common tool in the medical field, the domain general- ization problem becomes more important. Domain generalization is a deep-learned model’s ability to generalize to unseen sources. That is, to perform well on sources that were not seen during training. In image recognition and classification, especially in a medical setting, this is important because domain shifts can cause accuracy of medical image classification to drop significantly. Recent work in 2021 proposed a method named Stochastic Weight Av- eraging Densely (SWAD) that, through finding a flatter minima on the loss function during training, decreases the generalization gap between seen and unseen data. In this thesis, we show that the SWAD learning algorithm can increase the accuracy on out-of-domain chest X-ray data by 4.23% (P-value < 0.05) on average. In addition to this finding, an alteration to the SWAD algorithm is introduced that allows an increase in accuracy over the original algorithm on the chest X-ray dataset by 2.27% on average (2 tests P-value < 0.05, 2 tests P-value > 0.05).

Degree

thesis:*
Name thesis:degree_name
MS in Computer Engineering (M.S.C.P.)
Level thesis:degree_level
Thesis
Discipline thesis:degree_discipline
Computer Science and Engineering
Grantor
University of South Florida
Year dc:date
2024

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Weinhofer, Brandon M.
Contributors dc:contributor
  • Lawrence O. Hall, Ph.D.
  • Dmitry Goldgof, Ph.D.
  • Shaun Canavan, Ph.D.

Subjects

dc:subject × 6

Identifiers

dc:identifier.*
Repository record dc:identifier
https://digitalcommons.usf.edu/etd/10261

Chain of custody

source
Harvested from
University of South Florida
Base URL
digitalcommons.usf.edu/do/oai/
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
citation

Weinhofer, Brandon M.. Exploring the Use of Enhanced SWAD Towards Building Learned Models that Generalize Better to Unseen Sources. Thesis thesis, University of South Florida, 2024. https://digitalcommons.usf.edu/etd/10261