Design and Performance Evaluation of an Ultra-Low-Power Spiking Neural Network For Reservoir Computing

Abstract

This article discusses the architecture and performance analysis of an ultra-low-power spiking neural network (SNN) with reservoir computing (RC) to be used in an effective time-series predictor. The spiking neural networks that are energy efficient and event-driven communication are used in the RC framework to predict chaotic time-based data. A combination of spiking neurons and a fixed recurrent reservoir allows a massive reduction of computational complexity, allowing it to be used in low-power applications like edge computing and real-time monitoring systems. The experiment aims at forecasting the Mackey-Glass time series which is a popular benchmark to compare time prediction models. The model can efficiently learn non-linear dynamic trends of the time series by using a sparse spiking model and training the output layer only using pseudo-inverse regression. The results of the experiments show that the suggested system has a low root mean square error (RMSE) of 0.0426 with a moderate firing rate, thereby achieving a compromise between the accuracy of the predictions and the power efficiency. The system proposed is confirmed by the experiments based on simulation, which proves the validity and scaling ability of the proposed system in real-world applications in environmental sensing, financial forecasting, and biomedical signal processing. The paper does show that SNN-based reservoir computing models have the potential to create ultra-low-power AI systems with the ability to efficiently and accurately process complex temporal dynamics.