Pushing the limits of large-scale kernel computations
Abstract
Kernel methods are broadly used in scientific computing and machine learning. These techniques are based on fundamental linear algebra templates that can, in principle, be solved reliably. Even so, the enormous size of the linear algebra problems is an obstacle to employing kernel methods. At this point, there is a lack of effective algorithms for solving the large kernel problems that arisein practice. The goal of this project is to develop the computational tools needed to apply kernel methods to modern data sets.Thisproject will lead to new algorithms for solving large-scale kernel problems reliably, robustly, and efficiently. The technical approach will combine techniques from modern numerical linear algebra and numerical optimization. In particular, this project will develop novel randomized algorithms to overcome previous scaling challenges. The target is to solve spectral clustering and ridge regression problems with 109 data points on a desktop workstation.Kernel methods have many scientific applications with relevance for DoD priorities. Current applications include geospatial inference, molecular dynamics, quantum chemistry, quantum information science, and high-energy physics.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Mar 15, 2024
- Source ID
- N000142412223
Entities
People
- Joel Tropp
Organizations
- California Institute of Technology
- Office of Naval Research
- United States Navy