Abstract: Modern machine learning systems achieve impressive performance, but are largely built under a closed-world assumption: that the data distribution does not change from the distribution of the training set. Real environments are open, dynamic, and filled with unknown unknowns. In such settings, knowing when a model’s output is reliable is critical.
This talk focuses on out-of-distribution (OOD) detection, a key capability for safe and reliable AI. The first part presents a mathematical theory of OOD detection that places state-of-the-art methods, largely heuristically derived, within a unified variational information-theoretic framework [1]. The theory provides plausible assumptions behind existing approaches and predicts new OOD detectors that are simple to implement and outperform state-of-the-art methods.
The second part of the talk addresses an often overlooked problem in practical deployment of OOD detectors, that is, OOD detectors depend on parameters that must be tuned, and a “given OOD” dataset is required [2]. In practice, such given OOD data may be difficult to obtain. We formalize this problem and introduce a new tuning strategy that uses only the model’s training data and achieves similar or better performance compared to tuning on given OOD data, enabling robust and practical deployment.
The ideas will be illustrated across applications including automatic target recognition, cyber-security, large language models, and radio-frequency (RF) fingerprinting.
[1] https://arxiv.org/pdf/2506.14194
[2] https://arxiv.org/pdf/2602.05935
Bio: Dr. Ganesh Sundaramoorthi is Senior Research Fellow/Director of Research at RTX Technology Research Center, which is the research center for RTX (encompassing Raytheon, Collins Aerospace, and Pratt & Whitney) and also Adjunct Professor of ECE at Georgia Tech. His research is in machine learning, computer vision, and artificial intelligence (AI), e.g., robustness, explainability, acceleration, and low size, weight & power. Prior to his current position, he was on the faculty of KAUST, where he led a research group in computer vision. His PhD was from Georgia Tech and he did postdoctoral work at UCLA in computer vision. He has led a number of internal and external research programs in AI including IARPA, NGA, ARPA-E and AFRL. He was area chair for leading AI conferences (IEEE/CVF CVPR & ICCV). He has more than 60 publications in AI and nearly 50 patents and/or applications.