Discovery Through the Lens of AI: Scaling Anomaly Detection with Learned Representations

Tuesday 29 Sept 2026, 15:00 → 17:00 US/Central

WH11NE Sunrise

15:00 → 16:00 Discovery Through the Lens of AI: Scaling Anomaly Detection with Learned Representations

Learning begins when expectations fail. Many tasks, including discovery, validation, and monitoring, depend on testing data against a reference model, with implications for science, safety, and control. Yet many of the most relevant use cases across scientific and engineering domains are becoming harder to address, owing to growing data complexity, the absence of signal models, and the subtlety of the deviations we seek to detect. Recent advances in representation learning have made it possible to build rich, structured representations of data, allowing anomaly detection methods to scale to levels of complexity previously out of reach. In this talk, I will focus on testing for collective anomalies in LHC data without a prior signal model. I will show how contrastive representation learning yields powerful summary statistics for anomaly detection, how to account for uncertainties, and how to design sensitive detectors in these learned spaces. I will conclude with a broader perspective on applications beyond physics: as learned representations become the common language of modern AI, statistically principled anomaly detection in these spaces may become a key tool for understanding and monitoring the models themselves.

Speaker: Gaia Grosso (IAIFI, MIT)

Fermilab_2026.pdf

GMT20260929-200003_Recording_1672x878.mp4