Woojoo Na

I am a PhD student in Electrical and Computer Engineering at Northeastern University, advised by Professor Jennifer Dy. I trained as a pure mathematician at Oxford, and I approach machine learning through the lens of geometry and topology: many hard problems become tractable once you find the right space, the right shape of the feasible set, or the right path. I use that structure to design algorithms with guarantees, across several areas of machine learning:

• Large language model inference. Routing in mixture-of-experts models, where each token's route is a point in a simplex of expert patterns. When is a routing rule optimal, and how can every device stay balanced on every batch without retraining or dropping tokens? I study this in theory and on open MoE models with recorded user-conversation traffic.
• Optimal transport. Monotone, descent-guaranteed updates on the Bures–Wasserstein geometry of covariance matrices (ITSPACE, ICML 2026).
• Generative modeling. One-step generation by consistency training: which noise levels a fixed training budget should be spent on, and why a small loss can still leave a poor generator.

At Oxford, I did undergraduate research with Professor Christophe Petit, Professor Thomas Lukasiewicz, and Dr. Andrey Kormilitzin.

Email  /  CV  /  Google Scholar  /  GitHub

• 07/2026: Presented ITSPACE as a poster at ICML 2026 in Seoul. [arXiv]
• 05/2026: ITSPACE was accepted at ICML 2026! [OpenReview]
• 09/2024: Joined Northeastern University as a PhD student.
• 07/2023: My paper, "RACH-Space: Reconstructing Adaptive Convex Hull Space with Applications in Weak Supervision", is now available.
• 08/2021: Discharged from the Korean Army as Sergeant after completing 18 months of service.
• 02/2020: Joined the Korean Army to serve as FDC in artillery.
• 01/2020: I presented our work on cryptography at CMS-Ewha Womans University.
• 12/2019: We presented at 17th IMA International Conference on Cryptography and Coding.
• 07/2019: Graduated from the University of Oxford with a B.A. (Hons) in Mathematics.

A proximal majorization–minimization method for the exact Bures–Wasserstein (Gaussian W₂) objective on covariances, with closed-form updates in a square-root factorization. Each iteration is guaranteed to decrease the objective in exact arithmetic, with a certificate-gap bound under inexact polar computations; the updates preserve positive semidefiniteness and support rank-restricted factors. On real-world covariance-alignment benchmarks, ITSPACE reaches low-BW-gap solutions substantially faster than BW gradient descent and sample-OT baselines.

An algorithm for labeling unlabeled data in weakly supervised learning from incomplete, noisy information about the labels, based on reconstructing an adaptive convex hull. It reported state-of-the-art performance on the weak-supervision benchmark tasks considered at the time.

Undergraduate and master's-level research and course projects, kept here for the record.

Using group-theoretic properties of Cayley hash functions to construct trapdoor attacks on their parameters. Undergraduate research at the Oxford Mathematical Institute.

Studied the framework of group-equivariant CNNs, its applications and limitations, with experiments on MNIST.

An online platform for fake-news detection with neural models, aimed at small investors exposed to fake news.

Explored the relationship between transformer encoder word embeddings and conceptual-space geometry.

CNN and character-level word-embedding models for concept and relation extraction from electronic health records.

Knowledge graph completion for ontology reasoning.

Source code credit to Dr. Barron