Contrastive Learning for Interpretable Anomaly Detection at Collider Experiments
ORCA (Organized Representation via Contrastive learning for Anomaly detection) is a two-stage framework combining supervised contrastive learning with autoencoders for interpretable anomaly detection at collider experiments The method addresses two key problems in generic event-level anomaly detection: uninterpretable anomaly scores and strong correlation with energy scale and object multiplicity On simulated HL-LHC data, ORCA achieves significant gains in both breadth and depth of sensitivity t
Analysis
TL;DR
- ORCA (Organized Representation via Contrastive learning for Anomaly detection) is a two-stage framework combining supervised contrastive learning with autoencoders for interpretable anomaly detection at collider experiments
- The method addresses two key problems in generic event-level anomaly detection: uninterpretable anomaly scores and strong correlation with energy scale and object multiplicity
- On simulated HL-LHC data, ORCA achieves significant gains in both breadth and depth of sensitivity to new physics signals compared to baseline autoencoder architectures
- The contrastive embedding enables interpretability by allowing maximum-likelihood template fits to attribute anomalous events to known physics processes with quantified uncertainties
- The framework accurately recovers injected signal yields, including for signals excluded from training, and characterizes unknown signals by their resemblance to known template processes
Why It Matters
This work bridges machine learning and high-energy physics by making anomaly detection both more sensitive and interpretable—two properties that have historically been at odds in collider searches. For AI practitioners working in scientific domains, ORCA demonstrates how contrastive representation learning can add structural interpretability to otherwise black-box anomaly detection systems, offering a template for applying similar approaches in other fields where both detection performance and explainability are critical.
Technical Details
- Two-stage architecture: Stage 1 learns an embedding space via supervised contrastive learning across diverse physics processes; Stage 2 runs a standard autoencoder within that embedding space to generate event-level anomaly scores
- Interpretability mechanism: Known physics processes occupy distinct regions in the contrastive embedding space, enabling a maximum-likelihood template fit to attribute events in anomalous samples to specific template processes with quantified uncertainties
- Benchmark: Evaluated on a simulated dataset consistent with High-Luminosity Large Hadron Collider (HL-LHC) conditions, comparing against a baseline autoencoder architecture
- Signal recovery: The fit accurately recovers injected signal yields, including signals excluded from embedding training, and characterizes signals absent from the template library by identifying which known processes they most resemble
- Higher-dimensional advantage: The embedding geometry carries higher-dimensional physics information compared to standard one-dimensional output fits, enhancing downstream statistical analysis
Industry Insight
- The ORCA framework demonstrates a generalizable pattern: combining contrastive representation learning with reconstruction-based anomaly detection can simultaneously improve sensitivity and interpretability, a principle applicable beyond particle physics to any domain requiring both detection performance and explainability
- As collider experiments approach HL-LHC conditions, the ability to attribute anomalous events to known processes with quantified uncertainties could accelerate discovery pipelines by reducing the need for manual signal characterization
- Researchers should consider contrastive learning as a preprocessing step for anomaly detection systems where downstream statistical analysis depends on understanding what type of anomaly was detected, not just that an anomaly exists
Disclaimer: The above content is generated by AI and is for reference only.