Fully nonlinear neuromorphic computing with linear wave scattering

The increasing size of neural networks for deep learning applications and their energy consumption create a need for alternative neuromorphic approaches, for example, using optics. Current proposals and implementations rely on physical nonlinearities or optoelectronic conversion to realize the required nonlinear activation function. However, there are considerable challenges with these approaches related to power levels, control, energy efficiency and delays. Here we present a scheme for a neuromorphic system that relies on linear wave scattering and yet achieves nonlinear processing with high expressivity. The key idea is to encode the input in physical parameters that affect the scattering processes. Moreover, we show that gradients needed for training can be directly measured in scattering experiments. We propose an implementation using integrated photonics based on racetrack resonators, which achieves high connectivity with a minimal number of waveguide crossings. Our work introduces an easily implementable approach to neuromorphic computing that can be widely applied in existing state-of-the-art scalable platforms, such as optics, microwave and electrical circuits. As the energy consumption of neural networks continues to grow, different approaches to deep learning are needed. A neuromorphic method offering nonlinear computation based on linear wave scattering can be implemented using integrated photonics.