Publication Date
2024
Document Type
Dissertation
Committee Members
Michael Raymer, Ph.D. (Advisor); Thomas Wischgoll, Ph.D. (Committee Member); Cogan Shimizu, Ph.D. (Committee Member); Olga Mendoza-Schrock, Ph.D. (Committee Member); Vince Velten, Ph.D. (Committee Member); Mateen Rizki, Ph.D. (Committee Member)
Degree Name
Doctor of Philosophy (PhD)
Abstract
Deep neural networks have great representational power. However, most deep neural nets today optimize directly for performance on a single task defined only by labeled training data. This excludes potential sources of knowledge and ways of learning which could improve their performance, and address challenges, such as explainability, which are pressing to the field. We propose a framework for neural network architecture which generalizes it to a graph of many semantically-meaningful variables. We call it the Multi-Semantic-Stage Neural Network (MSSNN). An MSSNN models its domain as a web of conditional probabilities, i.e. a collection of inter-related tasks which can learn from each other. This way of conceptualizing a deep learning model manifests in several immediately useful capabilities. Such a model can easily exploit more and different kinds of labels in its training data. It can supplement its training data by incorporating expert knowledge, both in the structure of its graph and in the form of hard-coded relationships between variables. Graph structures such as multiple incident edges on a single variable automatically lead to semi-supervised learning ability. And finally, the MSSNN’s ability to sample from learned joint distributions allows us to construct novel explanations, directly tied to actual causes of the model’s behavior. We perform an initial demonstration of these capabilities by constructing and evaluating Multi-Semantic-Stage Neural Networks of several sizes for a collection of computer vision tasks.
Page Count
191
Department or Program
Department of Computer Science and Engineering
Year Degree Awarded
2024
Copyright
Copyright 2024, some rights reserved. My ETD may be copied and distributed only for non-commercial purposes and may not be modified. All use must give me credit as the original author.
Creative Commons License
This work is licensed under a Creative Commons Attribution-NonCommercial-No Derivative Works 4.0 International License.
