Physics, Computer Science, Math
-2nd Year Students
-ASU Online Barrett Honors Students (fully remote work)
-ASU Online Barrett Honors Students (fully remote work)
216
Tempe
Fully remote/Remote considered
Neural networks excel when trained on massive, high-quality datasets, allowing them to fit new data. In the Natural Sciences, however, data are inherently noisy, sparse, and expensive to acquire, fundamentally limiting the amount and quality of training data available. While the community has achieved notable successes with AI in cases involving exceptionally large and well-curated datasets–most famously protein structure prediction, recognized with a Nobel Prize–these high-visibility examples are exceptions rather than the rule. By contrast, the performance of AI on scientific imaging data–our primary window into biological processes and early disease detection–has been inconsistent. In one illustrative example from our own work currently under review at PNAS with am undergraduate as first author, we demonstrate that an entire class of widely used methods, many heavily reliant on AI, systematically reinforce incorrect models of motion. More broadly, it is understood in our community that many high-profile AI-based image reconstruction tools fail to generalize beyond the specific datasets provided by their authors, despite publication in flagship journals. At the root of this problem lies a fundamental property of microscopy data: it is noisy and sparse. Crucially, however, noise is not just a nuisance: it encodes valuable physical information. Our proposed approach will leverage known physical laws to generate probabilistic realizations of the data. This will then enable AI models to infer, with quantified uncertainty, the probable underlying “noise-free” image by a method known as simulation-based inference. This Physics-informed strategy departs sharply from prevailing Physics-free paradigms that rely on reassigning intensities in images pixel to pixel based on what was gathered from training datasets alone. As a result, the proposed project will offer a principled foundation for robust image reconstruction with the potential to underpin early disease diagnostics, for example by detecting subtle signatures of neurological disorders from retinal optical coherence tomography images.
Programming, and calculus (I-III) are pre-reqs.
Steve Pressé
Tempe; Fully remote/Remote considered