Dr. Ravi Shankar, assistant professor in the School of Foundations and Mathematics, has been awarded the National Science Foundation's prestigious CAREER Award for his research on optimization algorithms for deep learning. The five-year award provides $625,000 in funding and represents one of the NSF's most competitive early-career recognitions.
Dr. Shankar's funded research program, titled "Adaptive Curvature Estimation for Non-Convex Optimization in Large-Scale Machine Learning," investigates the theoretical foundations of why certain adaptive optimizers — particularly variants of Adam — generalize better than others in practice, despite theoretical results suggesting they should not.
"This is a known puzzle in the field," Dr. Shankar explained at a research colloquium held to celebrate the award. "Theoretical convergence bounds for Adam are worse than SGD in many settings, but practitioners consistently find Adam generalizes better on real models. Understanding why requires a new theoretical lens, and that is what we are building."
The research will develop new tools in stochastic differential equations and random matrix theory to analyze the implicit regularization effects of adaptive gradient methods. The work has potential applications to training efficiency (faster convergence with less compute) and safety (understanding the loss landscape geometry that models settle into during training).
Dr. Shankar joined Meridian as a founding faculty member in 2022 after completing his PhD at Stanford and a postdoc at the Flatiron Institute's Center for Computational Mathematics. He teaches MATH-430: Optimization for Machine Learning and MATH-490: Advanced Topics in Learning Theory.
"Ravi's work is exactly the kind of research that makes Meridian distinctive," said Dr. Ingrid Holmberg, chair of the School of Foundations and Mathematics. "It is theoretically deep and practically motivated. The NSF recognized what we have known since we recruited him."
The CAREER Award begins in July 2026.