Our Research Interests

Quantum Neural Networks and Quantum Machine Learning

By combining themes in quantum optics and neural networks, we have introduced a design of fully quantum neural networks, quantum systems that perform non-classical tasks, while not requiring the pre-existence of a quantum computer or annealer. We follow the paradigm of random recurrent reservoir networks and apply the created network to the task of distinguishing incoming optical states of a variety of types (e.g., squeezed thermal, photon added or subtracted, Bell states, or bound entangled states) as quantum entangled or separable. Being based on a random network, the scheme is compatible with a wide variety of optically active physical systems that don’t necessarily need a high degree of control during their fabrication (e.g., random networks of self-assembled quantum dots).

Topological Polaritons

The field of topology has proven how macroscopic properties of physical systems can result in exotic features at the boundaries between topologically distinct materials. Following earlier ideas from electronic topological insulators and photonics, we have introduced the theory of topological polaritons and topological indirect excitons . In these systems, density currents propagating in chiral edge states are protected from scattering with disorder, which is a clear advantage for exciton based optical circuits. The theoretical scheme makes use of the natural magnetic field sensitivity and spin-orbit coupling effects of excitons and has been experimentally implemented by our collaborators . In this realization exciton-polaritons are confined in a lattice of micropillars to create a bandstructure.We are further developing the concepts of topological polaritons, where we aim to use interactions to induce new bandstructures . In fact interactions allow more freedom in the design of bandstructures, allowing, for example, antichiral edge states that propagate in the same direction at opposite edges of a strip .

Polariton Information Processing Devices

Exciton-polaritons are highly nonlinear, fast responding, and have long coherence times, making them good candidates for optoelectronic information processing devices. We have designed universally complete and scalable schemes of information processing with exciton-polaritons accounting fully for losses and disorder. These designs have been based on mimicking biological systems, such as neurons to allow for robust signal propagation.

An alternative scheme can be based on polaritons in lattices, which exhibit local bistability when coherently driven. When accounting for spin-dependent coupling between lattice sites we show that a polariton lattice can behave as a class 4 cellular automaton (Conway’s life), which itself is a universally complete classical information processing system. This system also makes use of a range of solitons to carry information.

Quantum Optics in Weakly Nonlinear Systems&Optical Neural Networks

To generate quantum effects from an optical system one typically requires nonlinearity stronger than the system decay rate. For example, the well known photon blockade effect requires the interaction energy between two photons to exceed the linewidth. Unfortunately, most photonic systems (e.g., microcavities or photonic crystals) have short photon lifetimes. We have designed different mechanisms to circumvent this problem.

One mechanism, now known as unconventional blockade , is based on using quantum interferences, which may be very sensitive to even small amounts of nonlinearity. The mechanism was verified experimentally, where its main feature is the appearance of antibunching in the weakly nonlinear regime. We have also considered mechanisms of induced blockade using inverse four-wave mixing , which can lead to the generation of significant quantum entanglement in weakly nonlinear coupled mode systems

Neural networks exploit massive interconnectivity to become highly efficient at certain tasks, such as classification, and pattern recognition. While biological neurons may operate individually on millisecond time scales, their simultaneous connection to several thousands of other neurons allows a parallelization of tasks far beyond the capabilities of complementary metal-oxide-semiconductor (CMOS) logic. Naturally, this observation has motivated research into artificial neural networks, including hardware implementations. We have considered a scheme of optical neural networks making use of the interference of wave ensembles scattered by a controlled potential . A model based on the Schrödinger equation was used to illustrate the wide applicability of the scheme, which is compatible with a range of micron-sized solid-state implementations where a spatially varying potential is available. This includes nonlinear optical systems, systems that can be microstructured by optical lithography or exciton-polariton systems. As a demonstration, the recognition of noisy characters was achieved by the network. The advantage of the scheme is that the required weightings making up the structure of the network are naturally encoded in an optical potential, which itself is the result of a known superposition of waves. The engineering of more complex neural networks remains challenging in hardware systems, since it is not easy to control the weights of connections between artificial neurons, even if they can be calculated theoretically. More recently, with collaborators we have considered using exciton-polaritons for the implementation of reservoir neural networks . These networks are fully randomly connected, yet are still highly capable of particular tasks such as image recognition, speech recognition, and nonlinear time series prediction.