Overall, I am broadly interested in building reliable machine learning systems under imperfect data and distribution shift. My research centers on three closely related directions: (i) learning with imperfect or long-tailed data (e.g., limited, noisy, or imbalanced datasets), (ii) quantifying the unknown (e.g., uncertainty estimation and out-of-distribution detection), and (iii) adapting to novel domains and environments (e.g., domain adaptation, continual learning, and reinforcement learning).

A central challenge motivating my work is the imperfect nature of real-world data, particularly in scientific and healthcare domains. In applications such as medical imaging, biology, and behavioral science, datasets are often small, fine-grained, and highly imbalanced, making conventional machine learning methods difficult to apply reliably. My research therefore focuses on developing methods that learn robust representations from imperfect supervision while remaining reliable when encountering previously unseen data.

More broadly, I aim to develop machine learning systems that can recognize the unknown, reason about uncertainty, and continuously adapt to evolving environments. By addressing these challenges, my long-term goal is to move toward open-world machine learning systems that learn efficiently with minimal supervision while remaining robust beyond the training distribution.

My current work explores these research topics in medical imaging and animal behavior analysis, where reliable learning from limited and imperfect data is essential for real-world deployment.