High-speed time-series prediction using compact memristor circuits with adjustable dynamics

Abstract

Ta$_2$O$_5$ nonvolatile memristors are used as compact, traceable, and well-controllable dynamic reservoir computing layers to perform time-series prediction tasks. The strongly voltage-dependent switching speed is utilized for information processing. It enables the configuration of tailorable programming and forgetting times in response to the positive and negative driving voltage pulses. Benchmarking this framework on time-series prediction problems reveals that the configurable forgetting dynamics enables a high prediction accuracy using a rather small number of memristive input channels. The training is based either on optimizing the output layer using linear regression with fixed forgetting times, or on optimizing the forgetting times as well. In the first case, six memristive channels, while in the second, only two memristive channels are used to demonstrate excellent prediction accuracy for the benchmark tasks. This scheme allows for the tunability of the operating frequency over many orders of magnitude: by adjusting the input voltage levels, the information processing speed of the same memristive dynamic layer can be increased from the kHz to the MHz range while maintaining excellent prediction accuracy. These findings demonstrate the merits of memristor based dynamic networks in the analysis, prediction and recovery of fast temporal signals, approaching telecommunication data rates.