Photonic REservoir computing QUantum correlation Set ORacle

In the last decades, a new type of technology has been developed, which is based on quantum mechanics: the theory describing nature in its microscopic scale. Indeed, its counterintuitive laws have proven able to tackle problems that were considered out of reach and achieve remarkable advantages. For instance, quantum hardware can tackle in a few seconds computational problems that are solvable in thousands of years by the most powerful classical computers. Furthermore, quantum systems offer the tools to obtain the highest level of security in communications, robust against any kind of eavesdropper. Despite these premises, however, this field, especially from the experimental point of view, is still very young and many challenges are still open. For example, it is still an open question how to verify the correct functioning of such technologies and assess that they are based on quantum features and not just simulating them. PREQUrSOR goes in that direction and offers an experimental tool for detecting quantum phenomena in arbitrary processes. This will be done by combining the study of quantum mechanical foundations with machine learning, which is a branch of artificial intelligence aiming to make computers learn from given examples and then correctly deal with previously unseen data. More in detail, we will build a prototype of an artificial neural network, a learning model resembling the structure of a human brain, on quantum photonic hardware. This model will use the quantum behavior of photons, to classify trustworthy quantum devices from faulty ones. The main novelties of this project are two: on one hand, we will provide the first implementation of an artificial neural network on a photonic platform. Given the wide use of classical artificial neural networks, providing their quantum version will introduce a very versatile tool for tackling new and diverse challenges. However, this is challenging because the learning process of a neural network requires nonlinear effects, which can be obtained through the interaction of quantum systems with the environment. However, this can also lead to loosing part of their quantum characteristics. A solution to this apparent deadlock is given by a novel photonic device, the quantum memristor, developed by the University of Vienna, which displays a behavior that is similar to that of brain synapses, while preserving the quantum features of photons. The second interesting feature of this project is that we will not to tackle a classical problem and look for a quantum speed-up or enhanced accuracy in the solution. Indeed, the proposed verification is only feasible through a quantum apparatus, giving an immediate practical use of these results. On the contrary, looking for an advantage would have been a much more elusive scope , given the often unfair comparison between the early stage of quantum technologies and the maturity of the classical counterpart.

Modern machine learning plays a central role in everyday life, from medical diagnostics and scientific research to entertainment and online services. However, today's most powerful machine learning models consume enormous amounts of energy, making further scaling increasingly unsustainable. The main goal of this project was to explore more energy-efficient ways of performing machine learning using light and quantum technologies. Optical platforms, which process information using light instead of electrical currents, offer a promising alternative. Unlike conventional electronic systems, they do not rely on resistive elements that dissipate energy as heat. Many key operations can be carried out through light interference, making them reversible and therefore requiring less energy. This opens the door to computation below fundamental energy limits that constrain standard digital technologies. A major challenge for optical and quantum systems is achieving nonlinearity, a crucial ingredient for learning and decision-making. In this project, we systematically investigated how photonic quantum platforms can exhibit nonlinear behaviors and how powerful they can be for learning tasks. We studied both real-world problems and problems naturally suited to quantum technologies, as well as hybrid approaches combining classical and quantum processing. A key experimental ingredient was the use of tunable integrated photonic platforms, which offer high stability in a compact footprint. One key result of the project was the demonstration that there are learning tasks for which quantum photonic systems can achieve higher accuracy than classical approaches. This advantage was rigorously proven and published in Nature Photonics. A second major achievement was showing that complex learning dynamics can be obtained even in very small systems. In particular, we proposed and implemented the first neuromorphic (brain-inspired) architecture based on single photons and a photonic device with feedback, the quantum memristor. This system naturally provides both nonlinearity and memory, making it well suited for processing time-dependent data. Remarkably, a setup using only a single such element achieved performances comparable to much larger learning systems. We also developed alternative strategies to enhance learning capabilities in quantum systems, for example by re-injecting input data multiple times, a protocol known as data re-uploading. This approach enables complex transformations between inputs and outputs, providing sufficient flexibility to address image recognition tasks, including the analysis of satellite images. An ongoing follow-up effort investigates quantum machine learning for Earth observation, aiming to analyze satellite images directly in space and transmit only essential information back to Earth, reducing data and energy costs. Finally, we combined quantum machine learning with quantum simulation, showing how learning algorithms can be used to find the ground states of quantum systems, where the physical system encodes the solution of specific computational problems, such as factorization. This result was published in Physical Review Research.