Alexander Lvovsky

Abstract:

Optics is an exciting route for the next generation of computing hardware for machine learning, promising several orders of magnitude enhancement in both computational speed and energy efficiency. However, reaching the full capacity of an optical neural network necessitates the computing be implemented optically not only for inference, but also for training. The primary algorithm for network training is backpropagation, in which the calculation is performed in the order opposite to the information flow for inference. While straightforward in a digital computer, optical implementation of backpropagation has remained elusive, particularly because of the conflicting requirements for the optical element that implements the nonlinear activation function. In this work, we address this challenge for the first time with a surprisingly simple scheme, employing saturable absorbers for the role of activation units. Our approach is adaptable to various analog platforms and materials, and demonstrates the possibility of constructing neural networks entirely reliant on analog optical processes for both training and inference tasks.

Abstract:

We exploit polarization self-rotation (PSR) in atomic rubidium vapor to observe spontaneous symmetry breaking and bistability of polarization patterns. We pump the vapor cell with horizontally polarized light while the vertical polarization, which is initially in the vacuum state, is resonated in a ring cavity. Microscopic field fluctuations in this mode experience cumulative gain due to the compound action of amplification due to the self-rotation and feedback through the resonator, eventually acquiring a macroscopic magnitude akin to an optical parametric oscillator. The randomness of these fluctuations results in a bistable, random macroscopic polarization pattern at the output. We propose utilizing this mechanism to simulate an Ising-like interaction between multiple spatial modes and as a basis for a fully optical coherent Ising machine (CIM).

Abstract:

Diffractive optical neural networks (DONNs) have emerged as a promising optical hardware platform for ultra-fast and energy-efficient signal processing for machine learning tasks, particularly in computer vision. Previous experimental demonstrations of DONNs have only been performed using coherent light. However, many real-world DONN applications require consideration of the spatial coherence properties of the optical signals. Here, we study the role of spatial coherence in DONN operation and performance. We propose a numerical approach to efficiently simulate DONNs under incoherent and partially coherent input illumination and discuss the corresponding computational complexity. As a demonstration, we train and evaluate simulated DONNs on the MNIST dataset of handwritten digits to process light with varying spatial coherence.

Abstract:

A prerequisite to the successful development of quantum computers and simulators is precise understanding of the physical processes occurring therein, which can be achieved by measuring the quantum states that they produce. However, the resources required for traditional quantum state estimation scale exponentially with the system size, highlighting the need for alternative approaches. Here, we demonstrate an efficient method for reconstruction of significantly entangled multiqubit quantum states. Using a variational version of the matrix-product-state ansatz, we perform the tomography (in the pure-state approximation) of quantum states produced in a 20-qubit trapped-ion Ising-type quantum simulator, using the data acquired in only 27 bases, with 1000 measurements in each basis. We observe superior state-reconstruction quality and faster convergence compared to the methods based on neural-network quantum state representations: restricted Boltzmann machines and feed-forward neural networks with autoregressive architecture. Our results pave the way toward efficient experimental characterization of complex states produced by the quench dynamics of many-body quantum systems.

Abstract:

Quantum state tomography is an essential component of modern quantum technology. In application to continuous-variable harmonic-oscillator systems, such as the electromagnetic field, existing tomography methods typically reconstruct the state in discrete bases, and are hence limited to states with relatively low amplitudes and energies. Here, we overcome this limitation by utilizing a feed-forward neural network to obtain the density matrix directly in the continuous position basis. An important benefit of our approach is the ability to choose specific regions in the phase space for detailed reconstruction. This results in a relatively slow scaling of the amount of resources required for the reconstruction with the state amplitude, and hence allows us to dramatically increase the range of amplitudes accessible with our method.