Date of Award
2025
Document Type
Dissertation
Degree Name
Doctor of Philosophy (PhD)
Department
Computer Science
Committee Chair
Tathagatha Mukherjee
Committee Member
Sundar Christopher
Committee Member
Chaity Bannerjee
Committee Member
Manil Maskey
Committee Member
Jacob Hauenstein
Research Advisor
Tathagatha Mukherjee
Subject(s)
Machine learning, Quantitative research, Prior-Bias Congruence (PBC)
Abstract
My dissertation argues that effective machine learning arise from the synergistic integration of domain-specific priors and congruent inductive biases. I demonstrate this principle through explorations with diverse data types: considering the inherent spatiotemporal structure of radio frequency signal data, focusing on salient features and underlying distributions within geospatial data, and adapting to the unique vocabulary and statistical properties of specialized scientific corpora. These examples show that aligning model assumptions (biases) with domain knowledge (priors) enhances performance and efficiency. Building on this evidence, I propose an empirical framework, Prior-Bias Congruence (PBC), a conceptual structure aimed at guiding the development of models that systematically leverage domain knowledge. The framework advocates for more principled methods to identify priors and design congruent biases, aiming to create systems that learn effectively from data-driven learning while leveraging the wealth of expert knowledge.
Recommended Citation
Ramasubramanian, Muthukumaran, "Prior-bias congruence : case studies across heterogeneous domains towards a unified machine learning framework" (2025). Dissertations. 474.
https://louis.uah.edu/uah-dissertations/474