AF:Medium:Collaborative Research:Estimation, Learning, and Memory: The Quest for Statistically Optimal Algorithms

Stanford University, Stanford CA

Investigators

Abstract

The goal of this project is to develop new, efficient algorithms that extract as much information as is possible from a given quantity of data. In particular, this research aims to develop an understanding of how to leverage structure that is present in natural language settings, medical and genomic settings, and network- or graph-based settings. Many fundamental types of structure are encountered repeatedly in widely varying scientific and technological settings; our goal is to build on a recent body of work that focused on the simplest unstructured settings, and develop broadly applicable tools and insights to these diverse settings. A central component of this project is a close interaction and transfer of ideas, problems, and techniques, between the theory community, the machine learning community, and the broader set of data-centric researchers and practitioners. From a technical perspective, this research focuses on three fundamental types of structure: geometric structure, algebraic or low-rank structure, and the structure that is present in sequential data (such as natural language). For the first two types of structure, the research focus is on understanding the possibilities and limitations in the sparse data regime where the amount of data is comparable to, or sublinear in, the dimensionality of the data. In the third setting, the focus is on understanding the role of memory for learning and prediction tasks. Beyond the direct research goals of the project, the PIs are extensively involved in teaching and outreach, including designing UW?s new data sciences curriculum, and developing new courses on algorithms and foundational aspects of data sciences at Stanford.