A path to energy-efficient spiking delayed feedback reservoir computing system for brain-inspired neuromorphic processors

Abstract

Following the computation revolution in the field of machine learning, the reservoir computing system has shown its promising perspectives toward mimicking our mammalian brains, with comparable performance to other conventional neuromorphic computing systems. In this work, we proposed a spiking delayed feedback reservoir (S-DFR) computing system, which is embedded with the temporal encoding scheme, the Mackey-Glass (MG) nonlinear transfer function, and the dynamic delayed feedback loop. By adopting the temporal encoding scheme as the signal processing module, pre- and post-neuron signals are represented by the digitized pulse train with alterable time intervals. Experimental results demonstrate its rich dynamic behaviors with merely 206μW of power consumption; furthermore, the system robustness is studied and analyzed through the Monte-Carlo simulation. To the best of our knowledge, our proposed S-DFR computing system represents the first analog integrated circuit (IC) implementation of the time delay reservoir (TDR) computing system.