Photonic Quantum Memristor Networks

In the last decades, computer science has experienced two paradigm changes of very different kinds. The first was brought by machine learning, which is the ability of computers to learn from experience, namely from data. As such, this type of algorithms al ready constitutes an integral part of our everyday lives, ranging from personalized advertisements and e -mail filtering to data management and medical purposes. Machine learning is also constantly evolving, to deal with new applications and growing amounts of data. For example, novel algorithms mimic the structure of human brain, to cope with memory bottlenecks, and base their functioning on electronic elements like the memristor, whose behavior resembles that of neural synapses. The second outstanding breakthrough, instead, has been brought by quantum computation, that adopts the counterintuitive laws of the microscopic world to outperform the capabilities of classical computers. In this context, a most promising platform is constituted by photonics, that e xploits the quantum features of photons and through which quantum computational advantage was recently demonstrated for the first time. The goal of this interdisciplinary project is to combine these two worlds and benefit from their joint remarkable potentialities. In detail, it will bring to the first example of a photonic neuromorphic machine learning architecture, oriented to real -world tasks. As such, the project will overcome the apparent incompatibility of the linear laws of quantum mechanics with the non- linearity of learning processes, through a novel photonic element: the quantum memristor. This device, recently developed in a collaboration between the CNR of Milan (prof. Roberto Osellames group) and the University of Vienna (prof. Philip Walthers group), displays a non-linear behavior, through a controlled interaction of photons with the environment, while preserving their quantum features. Hence, by combining the complementary expertise in photonic quantum computing, integrated quantum photonics and quantum information theory of the partners of the project, this project aims at developing the first instance of a tunable photonic quantum memristor neural network, capable of executing programmable finite discrete mathematical transforms. Furthermore, this implementation will be based on integrated photonic circuits, that concentrate complex setups within millimetric spaces and, as such, it will be compact and scalable. The versatility of this nonlinear processor will be shown by demonstrating real -life quantum-enhanced applications reaching from speech recognition to image identification, accelerated via quantum reservoir computing architectures.

In the first part of our project, we theoretically and experimentally explored the nonlinearities we can achieve, through an effective encoding our input data into quantum systems and how we can enhance this nonlinearity by performing weak measurements and by implementing a feedback rule, to equip our model with short-team memory. The latter is required to implement type of artificial neural network, called reservoir computing. In our implementation, we use tunable integrated laser written circuits, implemented on borosilicate glass, by the Politecnico of Milan. The next phase of our project involves expanding our focus to more complex networks, utilizing multiple memristors and aiming for practical, real-world applications.