Event Actions
Active Measurement: Building a Better Instrument for Scientific Discovery
Abstract:
Artificial intelligence (AI) is transforming scientific discovery by enabling breakthroughs in how we collect, analyze, and interpret massive datasets. But as powerful as AI has become, its predictions remain imperfect: they can introduce biases, make critical errors, and often lack the statistical rigor required for scientific measurement. In this talk, I will present our ongoing work on active measurement, an approach that integrates AI with human feedback and principled statistical estimation. By embedding Monte Carlo techniques into the model development cycle, active measurement adaptively selects samples for expert validation and updates models when resources permit, to directly solve the measurement task. I will illustrate this framework through case studies in computer vision and remote sensing: monitoring bird migration at continental scales using weather radar, estimating species identity and population size from photographs, and similar applications in astronomy and remote sensing. These case studies highlight how active measurement can enable not only practical tools but also new scientific insights, pointing toward a more reliable and statistically grounded role for AI in discovery.
About the Speaker:
Subhransu Maji is a Professor in the Manning College of Information and Computer Sciences at the University of Massachusetts Amherst, where he co-directs the Computer Vision Laboratory. He received his Ph.D. in Computer Science from the University of California, Berkeley (2011), and a B.Tech. in Computer Science and Engineering from IIT Kanpur (2006). Prior to joining UMass, he was a Research Assistant Professor at TTI-Chicago. His research focuses on developing visual recognition algorithms that perceive fine-grained details and learn effectively from limited data. He is also interested in applications of AI that address societal needs and advance scientific discovery, with recent work in ecology, remote sensing, and other scientific domains. Maji has worked with several academic and industrial organizations, including the University of Amsterdam, AWS AI, Google, INRIA, Microsoft Research, Johns Hopkins University, and the University of Oxford. He is a long-term organizer of the Fine-Grained Visual Categorization (FGVC) workshop series and co-organizer of the recent CV4Science workshop. His work has been recognized with a CAREER Award from NSF, as well as Best Paper awards at AAAI 2024 and WACV 2015.