Physics for Neuromorphic Computing

Physical properties of quantum systems, such as superposition and entanglement, could allow them to exponentially accelerate some computations. Yet, quantum systems being noisy and their states fragile, execution of digital algorithms is out of reach of existing devices. On the contrary, brain computes efficiently with weak analog signals, close to the noise level, through dynamical processes. Hence the idea behind quantum neuromorphic computing is to take inspiration from the brain, and use quantum dynamics and analog signals to compute with quantum systems.



We have performed simulations and experiments to demonstrate quantum neuromorphic computing with superconducting circuits. Indeed, superconducting circuits allow to couple quantum oscillators in a parametric manner, using nonlinear elements based on Josephson junctions. Different interactions can be driven and adjusted through drive parameters, that can be trained to obtain desired outputs.



We first show that coupled quantum oscillators can transform the input data in a nonlinear manner which allows to classify it with a simple neural network architecture such as reservoir computing. Then we assess advantages of different variables that can be used as output neurons, i.e., occupation probabilities and emitted field quadratures. Finally, we compare different training methods, based on training Gaussian boson sampling distributions and variational methods.