“How much can we reduce scientific data without losing science?”

Abstract

Lossy compression of scientific data is necessary for many different domain sciences. In the past 5-6 years, we have witnessed a drastic diversification of use cases ranging from classic visualization to more advanced lossy data compression in communications or memory. We are also observing a blooming of lossy compression algorithms and, more importantly, continuous progress in compression quality, speed, and ratio. The high demand, diversity of use cases, and specific constraints and requirements created the need for environments to customize compression pipelines. The lossy compression methodology has also progressed drastically with the SDRBench benchmark, the Libpressio unified API, and the Zchecker error analysis tool. This talk will review recent progress in lossy compression for scientific data, including the new FZ project developing tools to help users customize their lossy compressors. I will also discuss challenging open research problems such as establishing bounds of scientific data lossy compressibility, homomorphic compression, On-chip compression, and compression for AI.

Biography

Franck Cappello is senior computer scientist and R&D lead at Argonne National Laboratory. From 2016 and with the support of ECP (US Exascale Computing Project), he started exploring lossy compression for scientific computing. This research produced the SZ lossy compressor, the Libpressio unifying API, the Z-checker tool to assess the nature of lossy compression errors and the SDRbench repository of reference scientific datasets. Franck is IEEE fellow and recipient of two prestigious R&D100 awards (2019 and 2021), the 2024 Europar achievement award, the 2022 HPDC achievement award, the 2018 IEEE TCPP Outstanding Service Award and the 2021 IEEE Transactions in Computers Award for Editorial Service and Excellence.

Zoom Link:  https://northwestern.zoom.us/j/97259800044?pwd=MXdVZElMSkduY2piSXhIK1B0RzJIZz09

Panopto:  https://northwestern.hosted.panopto.com/Panopto/Pages/Viewer.aspx?id=dd36e655-8ce3-4602-ab08-b0eb0105c426

Research Interests/Area

Scientific data analysis/transformation, resilience

Designing and Evaluating Human-AI Collaboration

Abstract

How can artificial intelligence systems most effectively assist humans in problem solving? This talk will offer a framework for addressing this question in high-stakes, real-world computer vision applications in medicine and misinformation. Specifically, this talk will discuss empirical results from two large-scale digital experiments on the dynamics of physician-machine partnerships in store-and-forward teledermatology diagnosis and AI assistance in deepfake detection. The findings in both experiments reveal that overall increases in task accuracy by humans with AI assistance come with unexpected trade-offs on other performance metrics.

Biography

Matt Groh is assistant professor at Northwestern University Kellogg School of Management in the Management and Organizations department and a core faculty member of Northwestern Institute on Complex Systems (NICO). His research examines human-AI collaboration with a focus on misinformation, medical diagnosis, and empathy. Before joining Northwestern, Matt worked as a data scientist at multiple startups, a non-profit, the World Bank, and DARPA and cofounded a startup. Matt has a BA from Middlebury College where he majored in economics and minored in Arabic and mathematics and an MA and PhD from MIT in Media Arts and Sciences.

Zoom Link: https://northwestern.zoom.us/j/91563101700?pwd=aVdXbWRMdTJSd3hEcnlhYnhsQnJ6QT09

Panopto:  https://northwestern.hosted.panopto.com/Panopto/Pages/Viewer.aspx?id=51a093fb-285b-4846-bf07-b0ee012bd2dd

Research Interests/Area

Human-AI Collaboration, Computational Social Science, Affective Computing, Synthetic Media

White-Box Computational Imaging: Measurements to Images to Insights 

Abstract

Computation and machine learning hold tremendous potential to improve the quality and capabilities of imaging methods used across science, medicine, engineering, and art. Despite their impressive performance on benchmark datasets, however, deep learning methods are known to behave unpredictably on some real-world data, which limits their trusted adoption in safety-critical domains. Accordingly, in this talk I will describe white-box, interpretable methods for photorealistic volumetric reconstruction that match or exceed the performance of black-box neural alternatives. I will also present recent theoretical results that guarantee correct and efficient reconstruction using our white-box approach in nonlinear computed tomography. 

Biography

Sara Fridovich-Keil is a postdoctoral fellow at Stanford University, where she works with Mert Pilanci and Gordon Wetzstein on foundations and applications of machine learning and signal processing in computational imaging. She is currently supported by an NSF Mathematical Sciences Postdoctoral Research Fellowship. Sara received her PhD in electrical engineering and computer sciences in May 2023 from UC Berkeley, where she was advised by Ben Recht and supported by an NSF GRFP fellowship. Sara received her BSE in electrical engineering from Princeton University in 2018, where she was advised by Peter Ramadge and supported, in part, by a Barry Goldwater Scholarship.

Zoom Link: https://northwestern.zoom.us/j/95206250523?pwd=T05wK0k0c0ROcklDTlVMektxV2Rndz09

Panopto:  https://northwestern.hosted.panopto.com/Panopto/Pages/Viewer.aspx?id=f3ac4dd3-f183-4b50-8163-b0ee012c8e6d

Research Interests/Area

Computational imaging; signal processing; machine learning; computer vision

CS Department Events

Argonne Visit

Other Events & Opportunities

Entrepreneurship Open House

Startup Matchmaking at The Garage

CS News

NU News