Scientific inference often requires reconstructing hidden physical systems from measurements that are incomplete, noisy, or indirect. Machine learning methods can help make these reconstructions possible, but they can also introduce dependence on modeling assumptions that are difficult to detect. The goal of my project is to study how these assumptions affect scientific conclusions and to develop tools for evaluating when a reconstruction is reliable. I will compare different modeling choices, analyze which results are stable under these choices, and test methods for separating information supported by the data from information introduced by the model. The broader aim of this project is to make machine-learning-based scientific inference more trustworthy and interpretable in settings where direct measurement is limited.

I am participating in this SuperUROP because I want to apply my knowledge from inference and information theory (6.7800), together with my background in physics, to scientific inverse problems such as black hole imaging. I am interested in inference problems and generative priors, and I hope this project will let me make a meaningful contribution to the field while developing research that could potentially lead to a publication.