We use auto-encoders in an anomaly detection setting to search for SUEP (soft unclustered energy patterns) and SVJ (semi-visible jets) signals in a background of QCD events.

In our journal paper we propose a family of Conditional Dual Auto-Encoders (CoDAEs) models that can learn multiple anomaly detection scores from raw images of particle collisions.

• There are two encodes: one with high capacity ($f_R$) to capture details in its large bottleneck, $Z$, and a smaller one ($f_m$) responsible to learn a discriminative 2-dimensional latent space, $Z_m$, that can be directly used for anomaly detection.
• The encoder $f_m$ is learned by conditioning (operation denoted by blue circles and paths in the figure) $Z$ on $Z_m$.
• Then, conditioning occurs multiple times at different resolutions of the decoder, $D$.

We can define multiple anomaly scores:

• From the latent space, $Z$: like the KL-divergence w.r.t. a prior, $p(Z)$.
• From the auxiliary bottleneck, $Z_m$, where each component can be considered as a score.
• Lastly, from the decoder's reconstructions: for example a score can be the reconstruction error.

The repository is organized as follows:

• ad/ contains the source code that defines layers, models, metrics, etc.
• weights/ contains pre-trained weights.
• data/ contains an example of the data used in our experiments.
• The notebooks codae.ipynb, categorical_codvae.ipynb, dirichlet_vae.ipynb, and qcd_or_what_model.ipynb show the training of the respective models (and evaluation of only the latest two.) supervised_cct.ipynb is used to train the supervised classifier.
• n_tracks.ipynb provides a comparison of our models against both physics-based and supervised baselines.
• tf-lite_convert.ipynb shows how to optimize (quantize) a CoDAE model, and measure its inference time.

Installation with virtual environment (otherwise open in, e.g., Google Colab):

1. Clone the repository: git clone https://github.com/Luca96/dark-autoencoders.git.
2. Change directory: cd dark-autoencoders\.
3. Create the virtual environment (named "venv"): python -m venv venv.
4. Activate it: venv/Scripts/activate (Windows) or venv/bin/activate (UNIX).
5. Install dependencies: pip install -r requirements.txt.
6. (optional) Install Jupyter notebook (or lab): pip install notebook or pip install jupyterlab.

Please consider citing our paper, if using any of the provided code and approach in your own research or project.

@article{anzalone2024triggering,
  doi = {10.1088/2632-2153/ad652b},
  url = {https://dx.doi.org/10.1088/2632-2153/ad652b},
  year = {2024},
  month = {sep},
  publisher = {IOP Publishing},
  volume = {5},
  number = {3},
  pages = {035064},
  author = {Luca Anzalone and Simranjit Singh Chhibra and Benedikt Maier and Nadezda Chernyavskaya and Maurizio Pierini},
  title = {Triggering dark showers with conditional dual auto-encoders},
  journal = {Machine Learning: Science and Technology},
}