ML study · task implementation
DeepFalcon: jet-image ML
Convolutional VAE reconstruction and graph classification of quark/gluon jets.
The problem
Jet images can be modeled as dense images or sparse detector-hit graphs.
My contribution
Implemented a convolutional VAE and a three-layer graph classifier over sparse point clouds.
Engineering decisions
- Compress images into a 256-dimensional latent representation.
- Build nearest-neighbor graphs from active detector pixels.
Results & evidence
- GNN: 70.9% test accuracy and 0.77 AUC in the documented 50,000-jet experiment (80/10/10 split).
- VAE: best documented validation loss of 315.3. Results describe these experiments, not a general benchmark.
