PELICAN

Overview

PELICAN is an innovative network designed for applications in particle physics, particularly for tasks involving 4-momentum inputs, such as classifying jet constituents and predicting momentum. Unlike traditional machine learning methods that function as black boxes, PELICAN incorporates Lorentz-invariant features in its architecture, ensuring that the system respects fundamental physics symmetries. This results in a model that not only maintains a compact size but also delivers outstanding performance across various tasks.

The design of PELICAN takes a significant step forward in sample efficiency and generalizability by using mathematical operations rooted in physical and geometric principles. This makes it an excellent choice for physicists and researchers looking for robust models without the drawbacks of conventional machine learning approaches.

Features

• Lorentz-Covariant Architecture: PELICAN employs a design that preserves Lorentz symmetry, allowing it to solve complex tasks while being physics-informed.

• Reduced Model Size: By embedding the necessary symmetries into its architecture, PELICAN significantly decreases the model size without compromising on performance.

• Sample Efficiency: The model's equivariant nature enhances its ability to learn effectively from smaller datasets, improving generalization and sample efficiency.

• User-Friendly Scripts: Two main scripts are provided for classification and momentum regression, which can be executed with no installation necessary, making it accessible to users.

• Flexible Dataset Handling: It supports a customizable dataset input mechanism, allowing for both single and multi-class classification tasks using straightforward command-line arguments.

• GPU and CPU Compatibility: PELICAN can be run on GPUs for optimized performance or switched to CPU mode, depending on user needs.

• Model Checkpointing: Users can load model checkpoints for inference or to continue training, ensuring flexibility in their experimentation process.

• Detailed Logging: The system can report per-minibatch statistics during training, helping users to monitor the learning process effectively.