Neuromorphic computation in multi-delay coupled models

Abstract

Neuromorphic computing provides a promising platform for processing high-dimensional noisy signals on dedicated hardware. Using design elements inspired by neurobiological findings and advances in machine learning methodology, delay-coupled systems have recently been developed in the field of neuromorphic computing. Delayed feedback connections enable such systems to generate a complex representation of injected input in the internal state of single nodes, which in our context refer to hardware components with nonlinear behavior and without any memory. In contrast to classical combinatorial circuits or feed-forward networks, this state is not distributed in space but in time. Hardware implementations with low hardware component counts are therefore particularly easy to design for delay-coupled systems. In this paper, we present an argument for using delay-coupled reservoirs using multiple feedback terms with different delays. We present a theoretical analysis of the resulting system, discuss surprising effects pertaining to the precise choice of delays, and provide a guideline for the optimal design of such systems.