Research
We study how machine learning systems can learn effective representations for diverse data domains and tasks—and how those representations can be understood.
Our research spans both representation learning algorithms and representation analysis. We develop learning methods and model architectures tailored to structured, multimodal, medical, and physical data, while using information-theoretic tools to investigate what learned representations encode and how they support generalization.
Research Questions
- How can learning algorithms acquire representations that are effective for a given data domain and task?
- How are learned representations structured, and how do they evolve during training?
- How can heterogeneous modalities such as tabular data, text, and time series be represented in a shared learning framework?
- How can information-theoretic analysis help explain representation quality and generalization?
Research Themes
Tabular Learning
We develop learning algorithms for tabular prediction and generation, with a particular interest in representation learning for heterogeneous and structured features.
Multimodal Foundation Models
We study foundation models that integrate structured data with other modalities, including tabular–text and tabular–time-series learning.
Information-Theoretic Analysis of Representation Learning
We use information theory as a primary analytical framework to understand how representations are formed, what information they preserve, and how representation structure relates to learning and generalization.
Representation Learning for Medical AI and Physical AI
We develop representation learning methods for data and systems in medical and physical environments, where observations are often heterogeneous, incomplete, multimodal, and task-dependent.
Research Perspective
Our goal is to develop representation learning principles that are useful across different domains rather than being tied to a single task. We are particularly interested in the interaction between how representations are learned and how they can be analyzed and understood.