Reservoir Computing in memristive nanowire networks

In order to overcome the limitations given by transistor-based systems working in the Von Neumann architecture, it is necessary to develop new technologies and computing paradigms. In this framework, brain-inspired structures allow to perform spatio-temporal correlated operations, typical of neural circuits: in particular, the Memristor is one of the new analog devices which lets this approach possible. Defined as the fourth circuit element, the memristor is a passive element which exhibits non-linearities in its dynamics, thanks to the change in the internal resistance state due to the rearrangement of the atomic structure. Here, the computational capabilities of self-organized memristive nanowire (NW) networks are investigated by simulations: thanks to its resistive switching and fading memory properties, the system is capable of mimicking human brain's synapses basic functions in processing external signals. In particular, the complex nonlinear dynamics of the network allows it to be exploited as a Reservoir, in the so-called Reservoir Computing (RC). This approach can offer efficient temporal processing of recurrent neural networks with a very low training cost, which makes it perfect for temporal data classification. In particular, in Reservoir Computing, a dynamic system (reservoir) which offers short-term memory is capable of mapping non-linearly an input in a higher dimensional space to emphasize and extrapolate spatio-temporal input correlations, in order to simplify the data classification. The RC architecture developed in this work is composed by three main building blocks. Firstly, a pre-process is required in order to transform the data in electrical signals. The stimuli are then applied to the NW network, which evolves in conductive paths thanks to the memristive processes that happen at the cross-point junctions between the nanowires. Lastly, the reservoir states are collected and used as input for a supervised learning algorithm, implemented in a readout function (one-layer neural network, regression model). In order to reduce the physical dimensions of the device without affecting the overall performance, the concept of virtual nodes has been exploited as well: in particular, the data processed in the n-th virtual node is affected by the previous virtual node states, meaning that it carries out information of the near history of the dynamical reservoir. Thanks to this model, it is possible to perform temporal data analysis and solve complex tasks, such as speech recognition: in particular, the task requires the recognition of spoken digits in the form of audio samples. In order to maintain a bio-inspired structure, the audio signals are pre-processed with the Lyon's Auditory Model, which allows to transform sound waveforms in voltage spike trains, by simulating the human's auditory system. At last, the model is optimized by considering different degrees of freedom, such as electrodes configuration, number of virtual nodes, input pre-processing, and readout function parameters.