Improving AI’s Ability to Detect New Particles: A Multi-Background Approach for Physics Discoveries
Imagine trying to find a never-before-seen type of subatomic particle among trillions of known ones. That’s the challenge particle physicists face when searching for new physics at the Large Hadron Collider. Now, CDS PhD student Lily H. Zhang, along with her colleagues at the Fermi National Accelerator Laboratory, the University of Wisconsin-Madison, and CDS, including CDS Faculty Fellow Aahlad Puli and CDS Associate Professor of Computer Science and Data Science Rajesh Ranganath, have developed an innovative approach to train artificial intelligence for this needle-in-a-haystack problem.
In a new paper titled “Robust anomaly detection for particle physics using multi-background representation learning,” published in Machine Learning: Science and Technology, Zhang and her co-authors propose a novel method that could revolutionize the search for new particles.
For decades, physicists have been working to expand our understanding beyond the Standard Model of particle physics. As Zhang explains, “The traditional approach that physicists have taken in the past has been to posit mathematical theories with certain hypotheses, and then run experiments to test their hypotheses.”
However, this hypothesis-driven approach has not been as fruitful as hoped in recent years. In response, the high-energy physics community has begun to explore more data-driven methods. “This switch to a data-driven way of trying to discover new physics is where anomaly detection comes in,” Zhang says. Recent efforts have focused on building highly expressive models with as few assumptions as possible, aiming to detect anything unexpected under current physics understanding.
Zhang and her team argue for a middle ground between these approaches. “It’s really great to be data-driven and try not to rely on too many assumptions,” she notes, “But there’s a lot of information that we know about physics that we should be taking advantage of in our anomaly detection approaches.” This perspective led to their novel approach, which builds smart representations for anomaly detection.
This work is motivated by empirically observed failures of generative models for anomaly detection. For instance, such models, when trained on images of animals and vehicles, fail to detect even house numbers as anomalous.
Previous work by Zhang, Ranganath and fellow NYU PhD Mark Goldstein, “Understanding Failures in Out-of-Distribution Detection with Deep Generative Models, concludes that estimation error is to blame. Moreover, even minimal estimation error can lead to such failures, an insight that prompted Zhang to explore moving beyond the use of generative models alone for anomaly detection.
In their new paper, Zhang, Puli, Ranganath, and their collaborators introduce a multi-background approach to anomaly detection. “Multi-background” here refers to using multiple types of known particle interactions as training data for the AI, rather than just one dominant type. Instead of teaching the AI to recognize only the most common events as “normal,” it learns from several different types of well-understood particle interactions. This broader knowledge helps the AI more accurately identify truly anomalous events by giving it a more comprehensive understanding of what “normal” looks like in particle physics data.
Zhang emphasizes that avoiding false positives is crucial “to help prevent false discoveries.” The new approach reduces false positives by enforcing a set of robustness constraints. The recipe for robustness in anomaly detection is informed by Zhang’s earlier work, Robustness to spurious correlations improves semantic out-of-distribution detection, which draws heavily on prior theoretical work on robustness by Puli, Zhang, and Ranganath.
The researchers applied their technique to simulated data from the Large Hadron Collider, demonstrating improved performance in detecting unusual top quark jets while also reducing false alarms.
While further research is needed, this “smart representation” method offers an exciting new path for harnessing AI to push the boundaries of particle physics. By teaching machines to better understand the known universe, we may be one step closer to discovering the unknown.
By Stephen Thomas
