Semi-Supervised Gaussian Mixture Variational Autoencoder for Pulse Shape Discrimination

We address the problem of pulse shape discrimination (PSD) for radiation sources characterization by leveraging a Gaussian mixture variational autoencoder (GMVAE). When using PSD to characterize radiation sources, the number of emission sources and types of pulses to be classified is usually known. Yet, the creation of labeled data can be challenging for some classes as it requires expensive expert annotation. In this context, GMVAE can learn the distinct features of pulses from only unlabeled data. We show that classification accuracy can be further enhanced by adopting a semi-supervised GMVAE with auxiliary loss functions when labeled data are available. The preliminary results on two datasets with different number of classes suggest superior performance of GMVAE compared to other classifiers such as Gaussian mixture model (GMM) for unsupervised and semi-supervised learning and random forest for supervised learning.