Research
How can AI learn meaningful representations from imperfect observations?
This question lies at the heart of our research. We study the fundamental principles of representation learning that enable AI systems to generalize across structured, multimodal, and physical environments. Rather than focusing on a single application domain, we seek learning principles that transfer across diverse data modalities and real-world intelligent systems.
Research Questions
Our research is driven by questions such as:
- How do informative representations emerge during learning?
- Why do some learned representations generalize well while others fail under distribution shifts?
- How can AI learn robust representations from incomplete, heterogeneous, and multimodal observations?
- How can representation learning support reliable intelligence in real-world and physical environments?
Research Themes
Representation Learning
We investigate learning principles that produce informative, transferable, and robust representations. Our work spans self-supervised learning, information-theoretic learning, representation analysis, and foundation models.
Learning from Imperfect Information
Real-world data are often incomplete, noisy, heterogeneous, and sparsely labeled. We develop machine learning methods that remain reliable under missing information, limited supervision, and distribution shifts.
Structured & Multimodal AI
Modern AI systems increasingly integrate structured data with language, vision, time-series, and sensor observations. We study representation learning that bridges these heterogeneous modalities, with particular interest in structured multimodal foundation models.
Representation Learning for Physical AI
Future AI systems must understand and interact with the physical world through incomplete and multimodal observations. We investigate representation learning for physical AI, including healthcare, robotics, and other intelligent systems operating in real-world environments.
Research Philosophy
Rather than designing algorithms for individual tasks, we seek fundamental learning principles that generalize across domains. We believe representation learning provides a unifying perspective for understanding modern AI systems—from structured data and multimodal foundation models to physical AI.
Our goal is not only to build better AI systems, but also to understand why they learn, when they generalize, and how meaningful representations can support robust intelligence in the real world.