YSPH Biostatistics Seminar- "Integrating Generative Models and Optimal Transport for Structures Data: Decoding Latent Structures in Biomedical Data”

ABSTRACT

Recent advances in generative model learning paradigms, such as score matching and denoising diffusion probabilistic models, have elevated image and video generation to unprecedented levels of precision. However, what underlies the success of these models? How scalable are they to biomedical domains, particularly to structured data such as graphs, multivariate time series, and point clouds? This talk introduces Energy-based models as a versatile framework for modeling structured data distributions, emphasizing their integration with score matching techniques and their recent connection to diffusion models. Finally, we explore how these approaches intersect with optimal transport, a classical mathematical problem, and discuss their potential applications in improving anomaly detection in multi-modal biomedical data and accelerating drug discovery.

Bio: Lorenzo Simone is a third-year Ph.D. student in Computer Science at the University of Pisa, specializing in the advancement of generative models for structured data, such as multivariate time series, graphs, and point clouds. His research focuses on biomedical applications, including synthetic data for electrophysiology, molecular property optimization for drug discovery, and the detection of cardiorespiratory diseases. During his Ph.D., he has been awarded two research grants: one for using inertial measurement units (IMUs) in early cardiorespiratory disease detection and post-operative recovery, and a European research grant titled 'Investigation in Zebrafish and Drosophila melanogaster Autistic Spectrum Disorder (ASD) Models of Social Behavior Impairment,' which funded his position as a visiting assistant in research under the supervision of Prof. Shuangge Ma at Yale.