TL;DR
This paper introduces QUAK (Quasi Anomalous Knowledge), a novel deep learning approach for anomaly detection in high-energy physics. QUAK enhances sensitivity by incorporating "signal priors" – educated guesses about the features of new physics – into an unsupervised learning framework, allowing for more effective searches even when these priors aren't perfectly accurate.
What Problem Does This Paper Address?
The search for new physics at particle colliders like the Large Hadron Collider (LHC) often relies on identifying deviations from known physics. Traditional deep learning anomaly detection methods typically operate without any prior knowledge of what a new signal might look like. While this "prior-free" approach is model-independent, it can significantly reduce the sensitivity of the search. Conversely, highly specific searches are sensitive but can miss unexpected phenomena. This paper addresses the challenge of improving anomaly detection sensitivity by intelligently incorporating prior knowledge without sacrificing the broad applicability of anomaly detection.
Methods
The core of QUAK lies in its use of semi-supervised learning and a novel concept called "QUAK space." Instead of a single anomaly detection model, QUAK trains multiple unsupervised networks (specifically, Normalizing Flow Variational Autoencoders - NF-VAEs) on different "priors." These priors can be a background sample and various hypothetical signal samples that capture expected features of new physics.
The "QUAK space" is then constructed as an N-dimensional space where each dimension represents the "loss" (how well an event is reconstructed) from one of these unsupervised networks. By analyzing events within this QUAK space, researchers can identify anomalies that are "signal-like" (low loss in signal-prior networks, high loss in background networks) while also distinguishing them from general anomalies (e.g., detector glitches) that might have high loss across all dimensions. This approach effectively embeds prior knowledge into the anomaly detection process, guiding the search towards regions of interest.
Datasets
- MNIST Dataset: Used for initial demonstration and validation of the QUAK algorithm's ability to separate digits, emulating a simplified anomaly detection scenario.
- LHC Olympics 2020 Dataset (BlackBox 1): A simulated dataset designed to mimic real LHC data with a hidden new physics signal. This dataset was used to test QUAK's performance in a more realistic high-energy physics context, specifically for detecting dijet resonances.
Key Insights and Ideas
- Signal Priors Enhance Sensitivity: Even imperfect "signal priors" can significantly boost anomaly detection sensitivity compared to prior-free methods.
- QUAK Space as a Discriminator: The multi-dimensional QUAK space, built from the losses of specialized unsupervised networks, provides a powerful way to discriminate between signal-like anomalies and general background noise.
- Robustness to Prior Inaccuracy: QUAK demonstrates a remarkable ability to maintain performance even when the chosen signal priors deviate from the true signal, making it more resilient to uncertainties in theoretical models.
- Complementary to Existing Methods: QUAK can be used in conjunction with other anomaly detection techniques, offering a flexible framework for new physics searches.
Key Results
- MNIST Performance: QUAK successfully separated digits, showing improved performance over standard autoencoders when using proxy signals.
- LHC Olympics (BlackBox 1): QUAK identified an injected signal in the BlackBox 1 dataset with a significance exceeding 5 standard deviations, demonstrating its effectiveness in a realistic high-energy physics scenario.
- Comparison to Supervised Methods: In several tests, QUAK approached or even surpassed the performance of fully supervised classifiers, especially when the test signal deviated from the training signal of the supervised model. This highlights QUAK's robustness.
Limitations and Future Work
- Choice of Priors: While QUAK is robust to inaccurate priors, the initial choice of signal priors still influences the search. The paper suggests systematic scanning of potential priors.
- Computational Cost: Training multiple unsupervised networks for QUAK space construction can be computationally intensive.
- Future Work: The authors suggest exploring more advanced normalizing flow techniques, integrating QUAK with other anomaly detection methods (like classification without labels), and applying QUAK to a broader range of physics problems beyond particle physics, such as gravitational wave modeling and astrophysical signature identification. They also propose using QUAK for measurements where signal models have significant uncertainties.
Summary updated