Bio: Eric Nalisnick is an assistant professor at Johns Hopkins University. His research interests span statistical machine learning and probabilistic modeling, with an emphasis on quantifying uncertainty in deep learning, human-AI collaboration, specifying prior knowledge, and detecting distribution shift. He previously was an assistant professor at the University of Amsterdam, a postdoctoral researcher at the University of Cambridge and a PhD student at the University of California, Irvine. Eric has also held research positions at DeepMind, Microsoft, Twitter, and Amazon. His papers have been recognized with selective oral presentations (ECCV 2024) and awards (AIStats 2023, AIStats 2024).

Abstract: Anomaly detectors perform essential tasks ranging from protecting an autonomous system from abnormal inputs to isolating interesting scientific signals that may lead to discovery. However, how do we know that the anomaly detector will be robust? The detector itself could be susceptible to failure if there is a change in the distribution of anomalies expected. And to make matters worse, we usually don't have an abundance of test-time anomalies that are labeled as such. In this talk, I will discuss my group's recent work on monitoring if our anomaly detector is still valid and adapting it to data shift, without supervision.

Bio: Jennifer Ngadiuba, a Wilson Fellow at Fermilab since 2021, specializes in searching for new physics in collider data and advancing AI applications in high-energy physics. After earning her Ph.D. from the University of Zurich, she contributed to CMS experiment studies, focusing on diboson resonances and jet substructure techniques. Starting when she was research fellow at CERN and Caltech, she pioneered deep learning for anomaly detection and fast machine learning on FPGAs for real-time systems in particle physics. Her work earned her the DOE AI4HEP award and AI2050 fellowship in 2023, recognizing her transformative contributions to experimental physics and machine learning.

Abstract: Anomaly detection techniques have been proposed as a way to mitigate the impact of model-specific assumptions when searching for new physics at the LHC. In this talk I will discuss how these techniques, when based on modern AI developments, could be utilized at different stages of data processing workflow, from real-time systems to offline analysis, and the impact they could have to revolutionize the current paradigms in the search for new physics.

Bio: Adji Bousso Dieng is an Assistant Professor of Computer Science at Princeton University where she leads the lab Vertaix on research at the intersection of artificial intelligence and the natural sciences. She is affiliated with the Chemical and Biological Engineering Department, the Princeton Materials Institute, the Princeton Quantum Initiative, the Andlinger Center for Energy and the Environment, and the High Meadows Environmental Institute (HMEI) at Princeton. She is also a Research Scientist at Google AI and the founder and President of the nonprofit The Africa I Know. She has been recently named an Early-Career Distinguished Presenter at the MRS Spring Meeting,  one of 10 African Scholars to watch in 2025 by The Africa Report, an Outstanding Recent Alumni by Columbia University's Grad School of Arts and Sciences, an AI2050 Early Career Fellow by Schmidt Sciences, and as the Annie T. Randall Innovator of 2022 for her research and advocacy by the American Statistical Association. She received her Ph.D. from Columbia University. Her doctoral work received many recognitions, including a Google Ph.D. Fellowship in Machine Learning, a rising star in Machine Learning nomination by the University of Maryland, and a Savage Award from the International Society for Bayesian Analysis, for her doctoral thesis. Dieng's research has been covered in media such as the New Scientist and TechXplore. She hails from Kaolack, Senegal. 

Abstract: This talk will cover the concepts, tools, and methods that make up Vendi Scoring, a new research direction focused on the concept of diversity. I’ll begin by introducing the Vendi Scores, a family of diversity metrics rooted in ecology and quantum mechanics, along with their extensions. Next, I’ll discuss algorithms for efficiently searching large materials databases and exploring complex energy landscapes, such as those found in molecular simulations, using the Vendi Scores. Finally, I’ll introduce the new concept of 'algorithmic microscopy,' which stems from Vendi Scoring, and describe the Vendiscope, the first algorithmic microscope designed to help scientists zoom in on large data collections for data-driven discovery.

Bio: Suhee Yoon is an AI Research Scientist at LG AI Research, specializing in AI Safety & Reliability with a focus on robustness under distribution shifts and efficient adaptation of large-scale foundation models. She holds a Master of Science in Industrial Engineering from Sungkyunkwan University. Her research spans various data domains, including computer vision, chemistry, and tabular data, aiming to develop AI models that are both reliable and adaptable to real-world challenges.

Abstract: Out-of-distribution (OOD) detection, which determines whether a given sample is part of the in-distribution (ID), has recently shown promising results through training with synthetic OOD datasets. Nonetheless, existing methods often produce outliers that are considerably distant from the ID, showing limited efficacy for capturing subtle distinctions between ID and OOD. To address these issues, we propose a novel framework, Semantic Outlier generation via Nuisance Awareness (SONA), which notably produces challenging outliers by directly leveraging pixel-space ID samples through diffusion models. Our approach incorporates SONA guidance, providing separate control over semantic and nuisance regions of ID samples. Thereby, the generated outliers achieve two crucial properties: (i) they present explicit semantic-discrepant information, while (ii) maintaining various levels of nuisance resemblance with ID. Furthermore, the improved OOD detector training with SONA outliers facilitates learning with a focus on semantic distinctions. Extensive experiments demonstrate the effectiveness of our framework, achieving an impressive AUROC of 88% on near-OOD datasets, surpassing the performance of baseline methods by a significant margin of approximately 6%.