neural end-to-end-optimised summary statistics
arxiv.org/abs/2203.05570

Leverages the shoulders of giants (jax and pyhf) to differentiate through a high-energy physics analysis workflow, including the construction of the frequentist profile likelihood.

If you're more of a video person, see this talk given by Nathan on the broader topic of differentiable programming in high-energy physics, which also covers neos.

Some things need to happen first. Click here for more info -- I wrote them up!

Do you want to chat about neos? Join us in Mattermost:

Please cite our newly released paper:

@article{neos,
    Author = {Nathan Simpson and Lukas Heinrich},
    Title = {neos: End-to-End-Optimised Summary Statistics for High Energy Physics},
    Year = {2022},
    Eprint = {arXiv:2203.05570},
    doi = {10.48550/arXiv.2203.05570},
    url = {https://doi.org/10.48550/arXiv.2203.05570}
}

In a python 3 environment, run the following:

pip install --upgrade pip setuptools wheel
pip install neos
pip install git+http://github.com/scikit-hep/pyhf.git@make_difffable_model_ctor

With this, you should be able to run the demo notebook demo.ipynb on your pc :)

This workflow is as follows:

• From a set of normal distributions with different means, we'll generate four blobs of (x,y) points, corresponding to a signal process, a nominal background process, and two variations of the background from varying the background distribution's mean up and down.
• We'll then feed these points into the previously defined neural network for each blob, and construct a histogram of the output using kernel density estimation. The difference between the two background variations is used as a systematic uncertainty on the nominal background.
• We can then leverage the magic of pyhf to construct an event-counting statistical model from the histogram yields.
• Finally, we calculate the p-value of a test between the nominal signal and background-only hypotheses. This uses the familiar profile likelihood-based test statistic.

This counts as one forward pass of the workflow -- we then optimize the neural network by gradient descent, backpropagating through the whole analysis!

A big thanks to the teams behind jax, fax, jaxopt and pyhf for their software and support.