XTRIPODS: Engaging Underserved Communities of Learners in Data Science (EUCLID): A Data-Science Partnership Between University of Chicago and University of California Merced

University Of California - Merced, Merced CA

Investigators

Abstract

This project establishes the Engaging Underserved Communities of Learners in Data Science (EUCLID) program, which is a partnership between the University of California, Merced and the University of Chicago, as part of the TRIPODS Phase II Institute for Foundations of Data Science. The EUCLID program leverages strengths of faculty from these two institutions to provide undergraduate and graduate students a training experience that prepares them for careers in academia, industry, and government. The EUCLID project aims to broaden participation through research, educational, and workforce development activities. The three main objectives of the EUCLID program are as follows: (1) train graduate students in research in machine learning and optimization; (2) develop undergraduate research projects; and (3) develop a linear algebra course as part of the undergraduate Data Science and Computing major at UC Merced. Recent work in machine learning has demonstrated that deep learning techniques can be used for signal recovery and image reconstruction. Typically, deep neural networks require a large training set, and a predictor function is learned by solving an optimization problem for some given loss function. In many physics-based applications, the observation operator is known. The framework of deep unrolling guides the design of neural network architectures that explicitly incorporates knowledge of the physics of the sensing model. For certain problem classes, deep unrolling learns representations of the data that reflect the constraints imposed on the architecture based on the sensing model. In this work, the team expands on this approach in three ways. First, the work goes beyond gradient descent to optimize the network parameters and incorporate quasi-Newton methods, which exploit previously computed iterates and gradients. Second, the team aims to investigate alternative loss functions that better reflect the noise distribution of the measurements. Third, the project explores optimization techniques for nonlinear physical models. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.