A Physical Reservoir Computing Model Based on Volatile Memristor for Temporal Signal Processing

Reservoir computing has emerged as a practical paradigm of implementing neural network algorithms on hardware for high-efficient computing. With the concept of reservoir computing, various electronic' dynamics can be harvested as computational resources, which has received considerable attention in recent years. Volatile memristor is an emerging memristive device that exhibiting interesting biomimetic behaviours such as short-term memory. Moreover, its conductance state can be varied by historical stimulation. In this work, a reservoir computing model using TiOx-based volatile memristor as processing core is proposed. The volatile memristor is measured and characterised, followed by using the discrete model to approximate the behaviours of the volatile memristor. Finally, a parallel volatile memristor reservoir computer is simulated based on the volatile memristor model. This model is evaluated by a waveform classification. The results (normalized root mean square error is 0.15 when using 10 volatile memristors) indicate the feasibility of using the physical behaviours of volatile memristor for constructing reservoir computers.