Speaker: Dr. Gabriella Contardo
Host: Dr. Daniela Huppenkothen
From Individual Outliers to Distribution Divergences: Different Flavors of Anomaly Detection in Physics
Event details from: Speaker: Dr. Gabriella Contardo
- Date
- 13 November 2024
- Time
- 11:00 – 12:00
- Room
- C4.174
"Anomaly" can take on different meanings across scientific domains, reflecting conceptual differences in what makes data 'unusual.' In astronomy, anomalies often imply individual outliers: rare objects that deviate from the rest of the data population. In contrast, particle physics focuses more on detecting distributional anomalies, where the distribution of the observed data diverges from theoretical expectations (e.g. standard model). This shifts the search from object-level towards distribution-level, more general properties of the data. While these approaches may seem distinct, they represent complementary facets of a broader challenge: leveraging automated methods for scientific discovery in increasingly complex datasets.
I will present two recent works that illustrate this anomaly-detection spectrum: first, a data-driven contextual anomaly search pipeline, developed to identify mid-infrared excess in Sun-like stars --potential tracers for extreme debris disks, planetary collisions, or (more speculatively) techno-signatures. Second, a novel method for local density anomaly detection, demonstrated on the LHC Olympics dataset: in this regime, we look for extremely small variations in the data distribution compared to a reference set. Finally, I'll discuss broader considerations and open problems in anomaly searches and automated, data-driven discovery in high-dimensional scientific data.