Differentiating signal from artefacts in cosmic ray detection: Applying Siamese spiking neural networks to CREDO experimental data

Abstract

The Cosmic Ray Extremely Distributed Observatory (CREDO) is an international research consortium aimed at observing high energy cosmic ray particles. The associated Android/iOS application enables the registration of muons with smartphone devices. The ubiquity of the CREDO infrastructure entails virtually no control over the detectors’ working conditions. In order to tag artefacts appearing in the CREDO database, we propose a Siamese spiking neural network (SNN) model, trained by optimizing Earth Mover’s Distance (EMD) between spike train outputs of the SNN. We first test the feasibility of our approach by training models on MNIST images converted into the spiking domain with novel conversion schemes. Then, on a binary classification problem of signal/artefact discrimination on CREDO images our model has achieved a class-balanced accuracy of 96.35%, close to the existing non-spiking solutions. Notably, the model has shown adaptability to input data properties in terms of the spiking network activity sparsity and prediction latency.