Deep Learning-Based Variational Autoencoder for Classification of Quantum and Classical States of Light

Authors

Mahesh Bhupati^*, Abhishek Mall^*, Anshuman Kumar, and Pankaj K. Jha

Abstract

Advancements in optical quantum technologies hinge on generation, manipulation, and characterization of light via photon statistics. Traditional classification requires efficient detectors and long measurement times for high-quality statistics. Here, we introduce a semi-supervised variational autoencoder (VAE) that maps photon-statistics features of single-photon-added coherent (SPACS), single-photon-added thermal (SPATS), and mixed states into a low-dimensional latent space for quasi-instantaneous classification at low photon counts. Our VAE remains robust against experimental losses—finite collection and quantum efficiencies, limited detectors—and leverages transfer learning to classify new datasets without retraining. This deep learning approach promises rapid, accurate classification of quantum light sources even under poor detection conditions.