Non-von Neumann computing architectures using integrated optical reservoirs

Microprocessors following the von Neumann architecture have become extremely powerful over the past decades, largely caused by reducing the size of individual devices and the introduction of new materials. New computational workloads for cognitive data analysis have become increasingly important. Neural networks, which are often used to efficiently solve these workloads, can be mapped into novel, non-von Neumann system architectures with co-integrated memory and processing units to improve the performance of the underlying algorithms.

Reservoir computing is an example of a recurrent neural network for dynamic data analysis, which has been implemented into hardware using optical, electrical, and mechanical systems. Using bulk optical and electro-optical components for delayed-feedback systems is a popular approach to realize reservoirs with several hundreds of nodes. More compact approaches based on silicon photonics with reduced number of nodes have recently been shown. Opportunities and challenges for integrated photonic reservoir computing will be discussed in this presentation, including the layout of the photonic circuits, the development of non-volatile synaptic elements in silicon photonics, and integrated nonlinear photonic building blocks.