[Paper Review] Novel deep learning methods for track reconstruction

Steven Farrell, Paolo Calafiura+12 more|arXiv (Cornell University)|Oct 14, 2018

Medical Imaging Techniques and ApplicationsMedicine44 citations

TL;DR

The paper presents two deep learning approaches for HL-LHC track reconstruction using space-point data: RNN-based track building and GNN-based hit/segment classification, showing strong performance on ACTS-simulated data and highlighting scalability advantages over image-based methods.

ABSTRACT

For the past year, the HEP.TrkX project has been investigating machine learning solutions to LHC particle track reconstruction problems. A variety of models were studied that drew inspiration from computer vision applications and operated on an image-like representation of tracking detector data. While these approaches have shown some promise, image-based methods face challenges in scaling up to realistic HL-LHC data due to high dimensionality and sparsity. In contrast, models that can operate on the spacepoint representation of track measurements ("hits") can exploit the structure of the data to solve tasks efficiently. In this paper we will show two sets of new deep learning models for reconstructing tracks using space-point data arranged as sequences or connected graphs. In the first set of models, Recurrent Neural Networks (RNNs) are used to extrapolate, build, and evaluate track candidates akin to Kalman Filter algorithms. Such models can express their own uncertainty when trained with an appropriate likelihood loss function. The second set of models use Graph Neural Networks (GNNs) for the tasks of hit classification and segment classification. These models read a graph of connected hits and compute features on the nodes and edges. They adaptively learn which hit connections are important and which are spurious. The models are scaleable with simple architecture and relatively few parameters. Results for all models will be presented on ACTS generic detector simulated data.

Motivation & Objective

• Motivate the use of deep learning for track reconstruction at HL-LHC where traditional combinatorial methods struggle with high occupancy.
• Propose RNN-based track building to extrapolate and evaluate track candidates similar to Kalman Filter operations.
• Propose Graph Neural Network-based hit and segment classification on a space-point graph to identify true connections and tracks.
• Evaluate proposed models on ACTS-simulated detector data to assess scalability and accuracy.

Proposed method

• Use space-point (hit) sequences with LSTM-based RNNs to extrapolate next hits and construct tracks, including a Gaussian-output variant that yields predictive uncertainty.
• Train the sequential hit predictor models with mean-squared-error or Gaussian log-likelihood losses for probabilistic predictions.
• Construct and evaluate graph representations of hits on adjacent detector layers and apply an EdgeNetwork/NodeNetwork-based GNN to classify hits and segments.
• Implement two GNN tasks: (i) binary hit classification to identify track hits; (ii) binary segment classification to differentiate true hit pairs.
• Report performance metrics such as purity, efficiency, and accuracy for GNN classifiers on constructed graphs.

Experimental results

Research questions

• RQ1 Can RNN-based models effectively extrapolate particle trajectories and build tracks from space-point sequences akin to Kalman Filter procedures?
• RQ2 Can Graph Neural Networks accurately classify hits and segments in a graph of space-point data to recover track candidates?
• RQ3 How do RNN and GNN approaches compare in scalability and performance on HL-LHC-like data?
• RQ4 What are the limitations and next steps to integrate these methods into full combinatorial track finding pipelines?

Key findings

• RNN-based track building achieves high hit-prediction accuracy (99.93% for simple and 99.98% for Gaussian models) in a low-occupancy setting.
• The Gaussian RNN provides uncertainty estimates via predicted covariances, with pull distributions broadly consistent with predictions.
• GNN hit classification achieved 99.2% purity, 97.9% efficiency, and 99.4% overall accuracy on a partially-labeled graph.
• GNN segment classification achieved 99.5% purity, 98.7% efficiency, and 99.5% overall accuracy on constructed segments.
• GNN methods demonstrate strong potential for scalable track reconstruction on structured hit data, deemed the most promising approach for HL-LHC conditions.