Reservoir computing (RC) is a machine learning framework initially conceived as an alternative to mitigate the expensive cost in training recurrent neural networks [1]. A RC system consists of a reservoir, a nonlinear dynamic system that maps the input in a high-dimensional space, and a readout layer, a single layer network that can be trained with simple methods, e.g. a linear regression. The fact that the reservoir does not require training, relaxes the hardware requirements for a physical realization, thus motivating the use of the RC as a powerful tool to exploit numerous physical systems as data processing units [2].

In this study, we use the RC approach to compute tasks with a chaotic oscillator, a modified version of the Murali-Lakhsmanan-Chua circuit [3]. More concretely, we included a Pt/HfO₂/TiN nonvolatile memristive device in the circuit. Such devices exhibit a physical phenomenon called resistive switching consisting in a nonvolatile change of resistance upon the application of voltage across its terminals. In the particular case of our device, the resistance increases (decreases) only after surpassing a positive (negative) threshold voltage level. When included in the circuit, the memristive device acts as a programmable resistor, providing a mechanism to tune the circuit's overall dynamics. The circuit is used as a reservoir, the first stage of data processing. The input data is masked in the amplitude of a periodic signal that induces the circuit to generate oscillations. Since the circuit has a single output port, the whole transient response is tracked with a sampling stage, where each sample constitutes a virtual reservoir output, and therefore a feature for training the readout layer. We demonstrate the feasibility of this scheme experimentally, with a physical realization of the circuit; only the training and inference of the readout layer are executed off-line. We tested the approach with two tasks, the computation of logic functions and pattern classification. The whole RC scheme is able to successfully classify the inputs due the nonlinearity of the circuit. Results indicate that the presence of chaotic sequences does not directly imply a disadvantage, and that they can contribute to the computation as well. Finally, we show that the nonvolatile memristive device acts as a knob to improve the classification of the inputs, when the computing performance appears to be suboptimal.