The Machine Learning Landscape of Top Taggers

Abstract: Based on the established task of identifying boosted, hadronically decaying top quarks, we compare a wide range of modern machine learning approaches. Unlike most established methods they rely on low-level input, for instance calorimeter output. While their network architectures are vastly different, their performance is comparatively similar. In general, we find that these new approaches are extremely powerful and great fun.

• The paper presents a comprehensive evaluation of ML methods for top quark tagging, achieving AUC values near 0.99 and exceptional background rejection.
• It compares image-based, four-momentum-based, and theory-inspired techniques to highlight performance nuances among different architectures.
• The study underscores the promise of integrating physics constraints with ML models to enhance real-time trigger efficiency and manage systematic uncertainties.

The Machine Learning Landscape of Top Taggers

The paper "The Machine Learning Landscape of Top Taggers" undertakes a detailed evaluation of various machine learning techniques applied to the identification and analysis of top quark jets, particularly focusing on boosted, hadronically decaying top quarks. This exploration is contextualized within the advanced experimental frontier presented by the Large Hadron Collider (LHC) at CERN, where the presence of strongly boosted tops is a quintessential feature of several physics processes.

Introduction to Top Quark Tagging

Top quarks are of significant interest due to their unique interactions with the Higgs boson and their distinctive feature of decaying before hadronization. This property makes the identification, or "tagging," of top quarks a sophisticated task, pivotal to accurate measurements and searches at the LHC. Traditional methods, such as jet algorithms and kinematic feature extraction, have been employed with ATLAS and CMS experiments. These methods generally utilize high-level observables derived from calorimetry and tracking information. However, the intersection of these traditional techniques with modern ML methods suggests new prospects for efficiency and performance.

Machine Learning Approaches Explored

This study compares a suite of machine learning methods that differ in their architecture and the nature of input data used:

1. Image-based approaches consider calorimeter outputs as two-dimensional images, analyzed using convolutional neural networks (CNNs) similar to those used in computer vision.
2. Four-momentum-based methods leverage the momentum information of jet constituents, analyzed using diverse architectures such as dense networks and recurrent neural networks.
3. Theory-inspired models incorporate additional physical considerations such as kinematic symmetries or infrared safety, and utilize ML architectures like the Lorentz Boost Networks which integrate domain-specific knowledge into the learning process.

Each of these approaches reveals a nearly comparable performance in top tagging, often reaching high levels of accuracy, AUC, and background rejection. Notably, despite the methodological differences, these ML models illustrate the significant potential of applying deep learning to low-level, LHC-specific data.

Key Results and Comparative Performance

The study provides a thorough investigation with specific performance metrics:

• The most effective models, including ParticleNet, ResNeXt, and TreeNN, achieve area-under-curve (AUC) values nearing 0.98-0.99, reflecting superior classification capabilities.
• Background rejection rates at a 30% signal efficiency consistently exceed 1/1000 in the best-performing models.
• Ensemble approaches further improve performance, demonstrating the coherence gain achievable by amalgamating outputs from multiple taggers.

These results underscore that, while various architectures yield comparably high performance, the choice of tagger may pivot more on secondary considerations such as training complexity or computational demand than on raw classification ability.

Discussion on Implications and Future Directions

The conclusions drawn from this research imply that contemporary ML techniques have effectively "solved" the problem of top tagging from a purely classification perspective using calorimeter data. However, future advances will likely focus on integrating additional experimental data such as tracking or vertexing, addressing systematic uncertainties in ML outputs, and optimizing these models for real-time trigger applications at experimental facilities like the LHC.

As ML models continue to evolve, a deeper integration of physics-inspired constraints and systematic uncertainty handling within ML models will be crucial for their adoption within the high-energy physics community. This will entail addressing challenges such as model robustness against experimental variations, managing training sample biases, and devising more detailed theory-guided benchmarks.

Ultimately, as this study reveals, the interplay between innovative machine learning methods and theoretical insights facilitates a comprehensive understanding and enhancement of top tagging techniques, driving forward the frontier of particle physics.