xai4hep

Code for:

[1] Farouk Mokhtar et. al., Do graph neural networks learn traditional jet substructure?, ML4PS @ NeurIPS 2022 arXiv:2211.09912
[2] Farouk Mokhtar et. al., Explaining machine‑learned particle‑flow reconstruction, ML4PS @ NeurIPS 2021 arXiv:2111.12840

Overview

XAI toolbox for interpreting state-of-the-art ML algorithms for high energy physics.

xai4hep provides necessary implementation of explainable AI (XAI) techniques for state-of-the-art graph neural networks (GNNs) developed for various tasks at the CERN LHC. Current models include: machine-learned particle flow (MLPF), and ParticleNet. The layerwise-relevance propagation (LRP) technique is implemented for such models, and additional XAI techniques are under development.

Explaining ParticleNet using LRP will produce the following edge-R-graphs.

Fig.1 - The jet constituents are represented as nodes in (eta, phi) space with interconnections as edges, whose intensities correspond to the connection's edge R score. Each node's intensity corresponds to the relative p_T of the corresponding particle. Constituents belonging to the three different CA subjets are shown in blue, red, and green in descending p_T order. We observe that by the last EdgeConv block the model learns to rely more on edge connections between the different subjets.

Explaining MLPF using LRP will produce the following R-maps.

Fig.2 - This figure constitutes averaged R-maps for elements associated to charged hadrons (top), and neutral hadrons (bottom). We see that charged hadrons use more neighbor information than neutral hadrons.

Setup

We recommend using the requirements.txt file then installing xai4hep as a module by running

pip install .

Other ways to setup,

1. If you have access to the kubernetes PRP Nautlius cluster, then refer to this gitlab repo for the setup https://gitlab.nrp-nautilus.io/fmokhtar/xai4hep

2. Using docker

docker build docker/