Research
Great are the works of the Lord;
they are pondered by all who delight in them...
He has caused his wonders to be remembered;
the Lord is gracious and compassionate.
(Psalm 111:2,4, NIV)
Research Interests
I’m broadly interested in machine learning interpretability, computational linguistics, and the theoretical foundations of artificial intelligence. I’m especially drawn to questions like: How do we make neural networks more transparent? and How do linguistic structure and semantics interact in multilingual NLP models?
You can view a summary of my research projects and scholarly output on my
ORCID profile.
Current Projects
Recommending Energy Efficiency Retrofits for Neighborhoods Using Machine Learning
June 2025—Present
Wise stewardship of resources is essential, particularly in the realm of building energy. Many buildings are far less energy efficient than they could be, resulting in excessive energy use and costs. While modeling tools like EnergyPlus™ and urban building energy models (UBEMs) can simulate building energy consumption, these remain only simulations. Our project seeks to go further: using simulation outputs alongside additional data as inputs to neural networks that generate retrofit recommendations at the neighborhood level.
Note: EnergyPlus is a trademark of the United States Department of Energy.
In collaboration with Professor Jorge Silveyra (Lafayette College) and Dr. Chetan Tiwari (Georgia State University)
Interpreting Regression Neural Networks with Linear Surrogates
April 2025–Present
In this project, I evaluate the reliability of linear surrogates for interpreting neural networks. Using a metric I call the lambda score, I measure how well linear models can approximate the predictions and representations of trained networks. While surrogates often achieve high correlation, I show that this does not imply faithful approximation—and in fact, the remaining unexplained variance may correspond to the network’s actual decision logic. This suggests that simple linear proxies can be misleading, especially when key nonlinear behaviors reside in low-variance regions of the input space.
Looking ahead, I’m interested in developing a related framework for classification tasks to explore whether the disconnect between fidelity and accuracy observed in regression also arises in classification settings. I also aim to characterize the fidelity–accuracy gap more precisely by studying the relationship between λ(f) and the R² between the surrogate and the ground truth—analyzing when and why high surrogate fidelity fails to preserve predictive performance.
Publications
Jackson Eshbaugh.
Fidelity Isn’t Accuracy: When Linearly Decodable Functions Fail to Match the Ground Truth.
arXiv preprint arXiv:2506.12176, June 2025.
📄 PDF🔗 arXiv
💻 Code
Detecting French Idioms Using Neural Machine Translation Techniques
February 2025–Present
Idiomatic expressions remain a major challenge in neural machine translation (NMT), often leading to errors in both statistical and modern NMT systems. In this project, I’m adapting techniques that have been successful in identifying idioms in English corpora and applying them to French data. This work, currently in its early stages, will become my combined honors thesis in French and Computer Science.
Upcoming Projects
FuncLearn: A Functional Programming Language for Machine Learning
Status: Early Design
Machine learning is used across many disciplines — but for those outside computer science, working with Python and TensorFlow can feel unnecessarily complex. FuncLearn aims to provide a simple, intuitive, English-like functional language that lowers the barrier to entry. The language will compile to TensorFlow-based Python code, allowing users to import datasets, chain models, and train networks using expressive, composable syntax — no ML expertise required.
Long-Term Interests / Ideas
I’m also interested in accessibility within programming languages — particularly the potential for localized keywords (
e.g., using native-language syntax like si instead of if) and rethinking syntax structures for right-to-left
languages. While this work is still conceptual, it reflects a broader interest in language-inclusive design.