Research Topics

(Updated Jun 2026)

Full publications available on Google Scholar, DBLP, arXiv.

Research Summary: My research studies fundamental and applied problems in computer vision and machine learning, with a focus on AI for health. We build intelligent systems that learn from complex data, reason about and interact with the real world, and drive meaningful impact on healthcare. Besides learning algorithms, we ask: how AI should represent knowledge across imaging, text, and multi-omics data; what principles ensure robustness, trustworthy, and clinical reliability; and how algorithmic advances can translate into real-world healthcare domains. Our research aims to answer these questions by advancing machine learning methods in close collaboration with biomedical and clinical science.

* indicates equal contribution.   ** indicates alphabetic author order.   ‡ indicates authors working closely with me.

  Efficient and Scalable World Foundation Models

We explore how to design general-purpose algorithms that enable world foundation models to align efficiently across tasks, modalities, and data scales. Our focus is on task-agnostic pretraining, scalable adaptation, and architectures that support transfer and compositionality. We aim to unify learning across heterogeneous domains while minimizing supervision and compute overhead. To achieve this, we develop methods that support dynamic task alignment, modular learning, and cross-domain generalization. These algorithmic advances lay the groundwork for building universal models applicable to science, medicine, and beyond.

  (Robust) Machine Learning for Imperfect Data

The development of machine learning models, particularly in the context of label scarcity, increasingly necessitates the collection of substantial annotated data. Moreover, massive data often display a long-tailed class distribution or subpopulation shifts, which consequently results in notable imbalance issues. To this end, there are several growing interests in training machine learning models jointly across imbalanced subpopulation distributions and limited annotations. We are developing novel algorithmic and computational approaches to ensure the efficiency and robustness of large machine learning models. Our applied research includes applications to healthcare, biomedical imaging, and cognitive neuroimaging.

  World Foundation Models for Biomedical Data

The development of medical world foundation models often requires massive and diverse biomedical data. To this end, we have developed various world foundation models for biomedical imaging data and explored novel applications of these models. We have also developed novel biomedical AI Agents that lead to the scalable and accurate predictive modeling, particularly for distribution shift problems.

  Machine Learning for Neural Systems and Biological Dynamics

We study machine learning methods for neural signals, brain connectivity, and biological state transitions. Neural and biological data often provide partial observations of high-dimensional systems, and computational models offer a way to infer structure from these measurements. EEG recordings are noisy and subject-specific, connectome data can be missing, and fixed-cell imaging captures endpoint populations rather than individual cell trajectories. Our work develops continuous-time generation, imputation-aware prediction, and geometry-constrained stochastic transport, with brain–computer interface (BCI) applications in data augmentation, subject adaptation, and neural decoding.

  Learning with Theoretical Guarantees

As machine learning methods have become ubiquitous in human decision-making, their reliability and interpretability have become important. This is particularly crucial in domains where decisions carry significant consequences, interpretable models can uncover crucial but unexpected patterns that complex models often obscure. We are currently studying provably interpretable modeling with theoretical guarantees. We are also exploring structured sparsity and attention in deep neural networks to enable interpretability.

  Learning with Multi-Modality Data

Multi-modality data is ubiquitous in healthcare and science applications. We are pursuing various techniques for modeling such multiple data, primarily using probabilistic graphical models and other statistical analyses. These tools are primarily used to facilitate biomedical research. We are developing various tools to effectively tackle real-world challenges associated with data heterogeneity. Of particular interest are novel methods that address robustness issues, such as confounding, as well as trustworthy computational approaches, with primary applications in healthcare, biomedical imaging, and cognitive neuroscience.