Scalable accelerated algorithms for exascale simulation and optimization-deep learning
Abstract
We propose to develop new scalable accelerated algorithms to enable exascale computing and dramatically increase the scale and size of solvable problems in the synergistic domains of (i) simulation and optimization of physical systems, and (ii) training deep neural networks. Towards this goal, we will devise new classes of accelerated first order methods with an eye towards applying these methods to general nonlinear systems arising from numerical simulation of physical processes and also to deep neural networks. While first order methods have lower complexity per iteration by orders of magnitude compared to second order or other information intensive schemes, they suffer from much slower convergence. The main research challenge lies in developing first order methods (FOMs) with improved converge rates and make them competitive with second order methods. In order to accelerate the convergence rate of FOMs, we propose to develop fast scalable preconditioners. The proposed algorithms will be implemented on next generation exascale platforms to enable numerical simulation of complex physical phenomena at unprecedented speed and accuracy, as well as the real time solution of ever larger and more complex deep learning applications.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jan 14, 2022
- Source ID
- FA95501910240
Entities
People
- Robert Freund
Organizations
- Air Force Office of Scientific Research
- Massachusetts Institute of Technology
- United States Air Force