Events

The Division of Health Informatics and Artificial Intelligence is excited to host Dr. Lav Varshney, Della Pietra Infinity Professor and Inaugural Director of the AI Innovation Institute at Stony Brook University.

Understanding complex system dynamics in settings such as public health, medicine, epidemiology, and the environment may require putting together several AI models and methods into compound AI systems that are complex configurations themselves. While individual AI models may excel across diverse domains—from protein analysis to climate prediction—such composition may enable enhanced performance, efficiency, interpretability, and inclusivity. Dr. Varshney presents several concrete approaches to demonstrate this paradigm. Equituning efficiently converts pretrained models into group-equivariant forms, improving compositional generalization. Platonic grounding optimizes multimodal AI models by processing multimodal tokens only in deeper layers where natural cross-modal alignment occurs, enhancing efficiency by reducing redundancy. Information lattice learning provides intrinsically interpretable knowledge discovery through hierarchical concept learning from universal priors. Efficient federated finetuning and federated digital twins enable bringing together data sources that may be restricted otherwise. He closes by discussing how these exemplary concrete approaches might be thought of in a unified calculus for building larger intelligent systems.