Electronic Theses and Dissertations
Date
2024
Document Type
Dissertation
Degree Name
Doctor of Philosophy
Department
Computer Science
Committee Chair
Deepak Venugopal
Committee Member
Vasile Rus
Committee Member
Xiaofei Zhang
Committee Member
Xiaolei Huang
Abstract
Neuro-symbolic models leverage the combined strengths of both symbolic AI and Deep Neural networks (DNNs), namely the interpretability of symbolic AI and the representation power of DNNs. These models are gaining significant attention, particularly in high-stakes domains where domain knowledge and interpretability play a key role. In this dissertation, we propose to develop neuro-symbolic AI based techniques to improve the scalability, verifiability, and equitability of these models. Specifically, we will focus on a high-stakes domain where the aforementioned properties are crucial, namely, AI-enabled education. In particular, Intelligent Tutoring Systems (ITSs) can make high-quality education available for a large and diverse population and we will ground our techniques in tasks related to discovering math problem-solving strategies from large-scale, real-world ITS data. To scale up, we will leverage symmetries/invariances in strategies to train models efficiently. To improve equitability, we will develop models that identify strategies for students with differing levels of mastery. To verify models, we will introduce a novel framework where we encode verifiable properties using first-order logic and learn a probabilistic model (Hybrid Markov Logic Network) to estimate uncertainty in representation learning. To reduce uncertainty in model representation, we will introduce a mixture model where each component in the mixture represents a uniquely parameterized HMLN and to compensate for covariate shift between training and test distribution, we will introduce the parameterization technique.
Library Comment
Dissertation or thesis originally submitted to ProQuest.
Notes
Embargoed until 05-25-2025
Recommended Citation
Shakya, Anup, "Neuro-Symbolic AI Techniques to Advance Model Verification, Scalability and Equity with Applications to Math Learning" (2024). Electronic Theses and Dissertations. 3673.
https://digitalcommons.memphis.edu/etd/3673
Comments
Data is provided by the student.