Classification and reconstruction of optical quantum states with deep neural networks

We apply deep-neural-network-based techniques to quantum state classification and reconstruction. demonstrate high accuracies reconstruction fidelities, even in the presence of noise with little data. Using optical states as examples, we first how convolutional neural networks (CNNs) can successfully classify several types distorted by, e.g., additive Gaussian or photon loss. further show that a CNN trained on noisy inputs learn identify most important regions data, which potentially reduce cost tomography by guiding adaptive data collection. Secondly, quantum-state density matrices using incorporate quantum-physics knowledge. The knowledge is implemented custom neural-network layers convert outputs from standard feedforward valid descriptions states. Any feed-forward architecture be adapted for (QST) our method. present demonstrations proposed [arXiv:2008.03240] QST technique conditional generative adversarial (QST-CGAN). motivate choice learnable loss function within an framework demonstrating QST-CGAN outperforms, across range scenarios, functions. For pure noise, able adapt reconstruct underlying state. reconstructs up two orders magnitude fewer iterative steps than maximum likelihood (iMLE) Further, both mixed randomly chosen points iMLE.