MORE | Fall 2020

A Novel Approach to Perform Rank-one Updates in Machine Learning

Data proliferation has made Machine Learning (ML) algorithms ubiquitous. However, the fact that ML algorithms can suffer from significant rounding errors and affect their output has received little attention. Moreover, techniques for avoiding rounding errors tend to be computationally expensive. In this research, we address both these issues by developing efficient roundoff error-free algorithms for solving the sequence of Systems of Linear Equations (SLEs) at the core of ML algorithms. Specifically, we take advantage of the fact that these SLEs are similar to each other to develop integer-preserving rank-one updates that avoid having to solve the SLEs from scratch.