Algorithms Design for Scalable Numerical Optimization
Dr. Aditya Devarakonda
, Wake Forest University
Abstract:
This talk will present communication-efficient design of stochastic optimization methods by combining design patterns from federated learning and s-step Krylov methods. We design a 2D, hybrid stochastic gradient descent algorithm, which attains a continuous performance tradeoff between federated SGD and s-step SGD. We perform computation and communication cost analysis highlighting the algorithmic tradeoffs, as well as, empirical performance on modern multi-core, multi-node hardware.
Speaker’s Bio:
Aditya is an Assistant Professor of Computer Science at Wake Forest University. His research interests are at the intersection of high performance computing, algorithms, and numerical optimization with emphasis on scalable algorithms design. He is currently on junior faculty leave from Wake Forest and is visiting ORNL during the 2024-2025 academic year.
