Online time-series forecasting using spiking reservoir

Abstract

IoT-based automated systems require efficient online time-series analysis and forecasting and there is a growing requirement to enable such processing at the low-cost constrained edge devices. Classical approaches such as Online Autoregressive Integrated Moving Average (Online ARIMA), Seasonal ARIMA (SARIMA) etc. and Artificial Neural Network (ANN) based techniques including Long-short Term Memory (LSTM) do not cater to this niche requirement due to their memory and computation power requirements. Neuromorphic computing and bio-plausible spiking neural networks, being both data and energy efficient, may offer a better solution. In this work, a novel spiking reservoir based network is proposed for online time series forecasting that relies on temporal spike encoding with a feedback driven online learning mechanism. The proposed network is capable of avoiding rapidly fading memory problem. The prediction accuracy of the network (tested on nine time-series datasets) outperforms conventional methods like SARIMA, Online ARIMA, Stacked LSTM, achieving up to 8% higher R2 score while using negligible buffer memory.