Reproduce “Particle Trajectory Representation Learning with Masked Point Modeling” on OpenResearch

About this paper

Sam Young, Yeon-jae Jwa, Kazuhiro Terao

Researchers at Stanford University and SLAC National Accelerator Laboratory introduced PoLAr-MAE, a self-supervised masked autoencoder designed for 3D point cloud data from Liquid Argon Time Projection Chambers (LArTPCs). This framework effectively learns physically meaningful representations of particle trajectories from unlabeled data and achieves comparable or superior semantic segmentation performance to supervised baselines using significantly less labeled data, reducing the reliance on extensive simulated datasets.

Key insights

• PoLAr-MAE learns to semantically distinguish particle types (e.g., tracks and showers) from unlabeled data, as evidenced by high F1-scores (0.994 for Track, 0.977 for Shower) in linear token classification.
• The model spontaneously learns to group points belonging to individual particle instances through its attention mechanism, demonstrating an emergent capability for instance segmentation without explicit supervision.
• Self-supervised pre-training significantly reduces the need for labeled data, enabling the model to achieve high performance on downstream tasks with considerably less supervision.

Summary by alphaXiv.