Pay Attention To Mean Fields For Point Cloud Generation

TL;DR

This paper introduces a novel attention-based generative model for point cloud data that maintains linear complexity, improves stability in adversarial training, and outperforms existing models on particle physics datasets.

Contribution

The authors propose a new point cloud generation method using attention aggregation with linear complexity and stabilized adversarial training, advancing particle physics data simulation.

Findings

Outperforms state-of-the-art GAN models on JetNet150 dataset.

Handles larger point clouds with up to 30 times more points.

Achieves stable training with feature matching loss.

Abstract

Collider data generation with machine learning has become increasingly popular in particle physics due to the high computational cost of conventional Monte Carlo simulations, particularly for future high-luminosity colliders. We propose a generative model for point clouds that employs an attention-based aggregation while preserving a linear computational complexity with respect to the number of points. The model is trained in an adversarial setup, ensuring input permutation equivariance and invariance for the generator and critic, respectively. To stabilize known unstable adversarial training, a feature matching loss is introduced. We evaluate the performance on two different datasets. The former is the top-quark \textsc{JetNet150} dataset, where the model outperforms the current state-of-the-art GAN-based model, despite having significantly fewer parameters. The latter is dataset 2 of the CaloChallenge, which comprises point clouds with up to $30\times$ more points compared to the first dataset. The model and its corresponding code are available at \url{https://github.com/kaechb/MDMA/tree/NeurIPS}.