Host: Ruzica Piskac
Title: Towards Creating Correct Numerics
Numeric application is an integral part of computer science. Real number computations are becoming increasingly important with the rapid development of machine learning, high-performance computing, and simulations. However, accurate and efficient numerical computations are still open research areas. A problem as simple as efficiently and correctly approximating elementary functions is still challenging and mainstream math libraries do not produce correct results.
This talk presents several challenges in numerical computation and proposes the RLibm project. This project builds efficient correctly rounded elementary functions for multiple representations. It makes a case for approximating the correctly rounded results of an elementary function rather than the real value of an elementary function. The talk then explores how this approach can be applied broadly and discuss future directions to take another step toward correct numerics.
Jay Lim is a Lecturer at Yale University. He received his PhD at Rutgers University under Santosh Nagarakatte. His research interest lies in making all low-level systems correct and secure. These systems include but are not limited to compilers, cryptographic algorithms, standard libraries, numerical applications, and ML. His dream is to see the day that all systems are free of bugs.