Research Symposium

BIO

Alex DeLise is a third-year student at Florida State University pursuing a B.S. in Applied and Computational Mathematics and a B.S. in Computational Science. His research background spans operations research, quantum computing, and scientific machine learning, with a current focus on generative modeling for inverse problems. He plans to pursue a Ph.D. in this or a related field, with the goal to develop more reliable and interperetable AI systems when solving real-world problems.

Alex began his research at the FAMU–FSU College of Engineering, working on optimization and data-driven modeling problems in operations research with Dr. Reza Abarzari and Dr. Arda Vanli. He later participated in two National Science Foundation Research Experiences for Undergraduates, where he studied quantum algorithms under Dr. James Ostrowski at the University of Tennessee, Knoxville, and scientific machine learning under Dr. Matthias Chung at Emory University.

He currently conducts undergraduate thesis research under Dr. Nick Dexter in the Department of Scientific Computing at Florida State University, where he studies active learning and generative modeling approaches to solving compressed sensing problems in imaging.

Active Learning for Conditional Generative Compressed Sensing

Abstract

Generative compressed sensing recovers high-dimensional signals from far fewer measurements than their ambient dimension by restricting solutions to the range of a learned generative model rather than the full signal space. For natural images, diffusion-based priors have demonstrated strong empirical performance in solving inverse problems, yet most existing approaches remain model agnostic, leaving measurement design decoupled from the structure of the learned image manifold and limiting efficiency in extremely undersampled regimes. In this work, we develop an active learning strategy that leverages Christoffel sampling of measurements tailored to a conditional diffusion model of natural images. Conditioning information, such as scene or picture acquisition metadata, is incorporated both in constructing the sampling distribution and during reconstruction, narrowing the set of plausible images and concentrating measurements on the most informative frequencies. This targeted sensing approach yields more accurate and stable reconstructions than uniform and variable density sampling when only a small fraction of coefficients are observed. Our results show that combining conditional generative priors with active measurement design can substantially reduce sampling requirements for natural image recovery.

Keywords: Machine Learning, Generative AI, Inverse Problems, Compressed Sensing