Inkee Jung

About Me

Welcome! a PhD candidate at Boston University working in geometry and topology, with expertise in machine learning and applied research.

I'm a member of the Geometry and Physics research group in BU.

My research advisor is Professor Siu-Cheong Lau

My undergraduate studies were in mathematics department at the Yonsei University in South Korea.

Research

I bridge geometry and topology with practical AI applications. I impose geometric and topological structure on datasets(Logifold) to make machine learning systems more interpretable and accessible through the lens of geometry and topology. My work includes developing a geometric framework for untrustworthy samples, including adversarial attacks, and a featurization method for Topological Data Analysis.

Publication & Preprints

Inkee Jung and Siu-Cheong Lau. (2026). Laws of Learning Dynamics and the Core of Learners, arxiv

Github

Inkee Jung, Wonwoo Kang, and Heehyun Park. (2025). Persistent Laplacian Diagrams, arxiv
Jung, I., & Lau, S.-C. (2025). A Logifold Structure for Measure Space. Axioms, 14(8), 599. https://doi.org/10.3390/axioms14080599
I. Jung and S. -C. Lau, "Logifold: A Geometrical Foundation of Ensemble Machine Learning," 2024 4th International Conference on Electrical, Computer, Communications and Mechatronics Engineering (ICECCME), Male, Maldives, 2024, pp. 1-6, doi: 10.1109/ICECCME62383.2024.10796884.

Talks

Teaching

2020 Fall - Present: Boston University
Teaching Assistant for Calculus I & II, Linear Algebra, Complex Variables, and Elementary Statistics.
Led problem-solving sessions and discussions for over 400 students across 8 semesters.
Summer 2023: Boston University, CAS MA 581 - Probability
Instructor for a summer course covering probability theory, Bayes' rule, random variables, distributions, moment-generating functions, the law of large numbers, and the central limit theorem.
Designed assignments and provided feedback throughout the course.
[Syllabus]
2020 Spring, Summer: Yonsei University
Teaching Assistant for Calculus I & II.
Grading and leading discussion sessions.

Selected Projects

Laws of Learning Dynamics and the Core of Learners - Spring 2025 - Spring 2026
Project on a structure of learning using logifold geometry and ensemble entropy to formalize a “core vs out-of-core” decomposition of data, then leveraging it for both when/why predictions are trustworthy and robustness under adversarial perturbations.
We use ensemble entropy (disagreement) to split inputs into reliable core vs suspicious out-of-core, then route OoC to specialists (IMM: Immunization Mechanism), and evaluate adversarial robustness. We generated transfer-based AutoAttack (L2) samples (PyTorch/RobustBench) and APGD samples (TF2/ART). We released notebooks covering attack generation, TF2 training (ResNet/VGG), sample generation, and the main evaluation pipeline in the following GitHub
We achieved improvements under strong adversarial attacks on CIFAR-10(APGD) from 21.48% (baseline) to 90.93% (IMM); on polluted environments (union of clean and perturbed samples), from 64.35% (U1 ensemble) to 87.48%(IMM). In non-clean setting, total entropy drops sharply from U1 to IMM, indicating much higher reliability of IMM predictions.
Predicting Aviation Accident Severity - The Erdős Institute - Summer 2025
GitHub | Video | Slides
Led data extraction and cleaning by converting .mdb files to .csv and integrating event, aircraft, and sub-aircraft tables into a unified dataset. Engineered and imputed features for classification and regression tasks. Built ensemble models (XGBoost, Random Forest, Extra Trees, Histogram Gradient Boosting) with grid search and cross-validation. Achieved an F1 score of 0.459 (Extra Trees) in aircraft damage classification, significantly above the 0.316 baseline (majority class).
Persistent Laplacian Diagrams - Spring 2025 - Present
Developing a featurization method for the Persistent Laplacian analogous to the persistence image for persistent homology, and defining signatures of the Persistent Laplacian together with a stability analysis for the resulting Persistent Laplacian diagrams.
Constructed an explicit pair of non-isomorphic graphs that are indistinguishable by persistent homology or classical graph Laplacian spectra, but distinguishable by our Persistent Laplacian based signatures, demonstrating their additional discriminative power.
Currently extending this framework to (bio)molecular dataset (e.g., QM7) for comprehensive evaluation of the method's practical utility in real-world applications.