Speech recognition through physical reservoir computing with neuromorphic nanowire networks

The hardware implementation of the reservoir computing paradigm represents a key aspect for taking into advantage of neuromorphic data processing. In this context, self-organised nanonetworks represent a versatile and scalable computational substrate for multiple tasks by exploiting the emerging collective behaviour of the system arising from complexity. The emerging behaviour allows spatio-temporal processing of multiple input signals and relies on the nonlinear interaction in between a multitude of nanoscale memristive elements. By means of a physics-based grid-graph modeling, we report on the implementation of reservoir computing for a speech recognition task in a memristive nanonetwork based on nanowires (NWs) acting as a physical reservoir. Besides analysing the pre-processing step for the transduction of the audio samples in electrical stimuli to be applied to the physical reservoir, we analyse the effect of the network size and the adoption of virtual nodes on computing performances. Results show that memristive nanonetworks allow in materia implementation of reservoir computing for the realisation of brain-inspired neuromorphic systems with reduced training cost.