QUAK

QUasi Anomalous Knowledge for Anomaly Detection and Tagging in High Energy Physics

Quasi Anomalous Knowledge: Searching for new physics with embedded knowledge

This repository is the official implementation of Quasi Anomalous Knowledge: Searching for new physics with embedded knowledge.

Requirements

I used conda to manage my dependencies.

To install requirements:

conda env create -f environment.yml

Datasets

QUAK used LHC Olympics dataset curated by Kasieczka, Gregor; Nachman, Benjamin; Shih, David, please cite https://zenodo.org/record/4536624 To read more about these datasets, please read LHC Olympics Community White Paper: https://arxiv.org/abs/2101.08320

We privately generated samples based on LHC Olympics dataset, the procedure of which is outlined in our paper: https://arxiv.org/abs/2011.03550 For training and evaluation, we applied pre-processing procedure which leaves each event with jet masses and substructure variables.

QUAK method is not limited to physics dataset; It can be applied to any environment where having vague knowledge of anomaly could help with the detection. To test QUAK in a different setting, we tested it on MNIST dataset (http://yann.lecun.com/exdb/mnist/).

Training

To train the model(s) in the paper, run this command:

python train_script.py

Evaluation

To evaluate QUAK performance on LHC Olympics black box dataset, run:

python eval.py --model-file mymodel.pth --benchmark imagenet

To evaluate QUAK performance on MNIST dataset, run:

Citation

We are preparing a journal submission, in the meantime, please cite our paper from arxiv:

@article{Park:2020pak, author = "Park, Sang Eon and Rankin, Dylan and Udrescu, Silviu-Marian and Yunus, Mikaeel and Harris, Philip", title = "{Quasi Anomalous Knowledge: Searching for new physics with embedded knowledge}", eprint = "2011.03550", archivePrefix = "arXiv", primaryClass = "hep-ph", month = "11", year = "2020" }

Pre-trained Models

You can download pretrained models here:

My awesome model trained on ImageNet using parameters x,y,z.
My awesome model trained on ImageNet using parameters x,y,z.

Results

Our model achieves the following performance on :

Image Classification on ImageNet

Model name	Top 1 Accuracy	Top 5 Accuracy
My awesome model	85%	95%

SangeonPark/QUAK