/GSoCSubmissionDeepfalcon

ML4SCI Deepfalcon

Primary LanguageJupyter Notebook

Project Overview

Below is a summary of each Task:

1. Variational Autoencoder with KL Divergence Loss

  • Utilized a variational autoencoder (VAE) architecture with KL divergence loss to reconstruct data.

2. Jets as Graphs

  • Converted images to point cloud format.
  • Generated edges using k-nearest neighbors (KNN) graph.
  • Classified graphs using a graph convolutional network (GCN).

3. Diffusion Model for Fast Detector Simulation

  • Implemented a diffusion model with a Time Conditioned UNet and DDPM scheduler to generate images.

4. Graph Transformer for Fast Detector Simulation

  • Utilized a vision transformer for image classification of quarks and gluons.
  • Used convolutional upsampling as a decoder to generate images.

5. Optimal Transport for HEP

  • Employed MNIST digits 1 and 9 for the dataset.
  • Implemented an autoencoder with latent dimensions mapping to a normal distribution for reconstructing digit images.
  • Utilized Sinkhorn loss for optimal transport.
  • Extended the model for quarks/gluon dataset image generation.