Date of Award:
8-2024
Document Type:
Thesis
Degree Name:
Master of Science (MS)
Department:
Electrical and Computer Engineering
Committee Chair(s)
Jacob Gunther
Committee
Jacob Gunther
Committee
Todd K. Moon
Committee
Kevin Moon
Abstract
The Codec Classifier is a low-computation, low-memory tree ensemble method that dramatically improves feasibility of image classification on resource-constrained edge devices. It achieves advantages over other tree ensemble methods due the separation of encoder and decoder tasks in the classifier. The encoder partitions feature space, and the decoder labels the regions in the partition. This functional separation of tasks enables the encoder design (partitioning) to be guided by maximizing the mutual information (MI) between class labels and the features (i.e. the encoded representation of the data) without regard to the error performance of the classifier. Experiments show maximizing MI leads to sequential partitioning of feature space that is more efficient than additive classifier models such as AdaBoost. The design results in gray-coded partitions in which adjacent regions are addressed by codewords that differ in only one bit. This novelty affords classification insights not available with other methods. The method is applied to binary classification (face detection) and multiclass classification (MNIST digits) problems.
Checksum
7c4164de4f3b95ed7155f6f36b495432
Recommended Citation
Beckwith, AJ, "Ensemble Machine Learning at the Edge Using the Codec Classifier Structure and Weak Learners Guided by Mutual Information" (2024). All Graduate Theses and Dissertations, Fall 2023 to Present. 227.
https://digitalcommons.usu.edu/etd2023/227
Included in
Copyright for this work is retained by the student. If you have any questions regarding the inclusion of this work in the Digital Commons, please email us at .