Multielectrode characterization of neuromorphic nanowire networks

The recent spread of AI and machine learning, as well as computational intensive tasks, launches the challenge of overcoming traditional Von Neumann architectures. Traditional computing architectures are reaching their limit, and the need for alternative computational paradigms arises. One of the most promising solution is represented by neuromorphic systems, computational architectures that mimic the behavior of biological neural networks, by processing and storing informations in the same framework. For this application the memristor is one of the most promising candidate, being an analog device that inherently mimics a biological synapse, exhibiting intrinsic memory and resistive switching features. The aim of this work is to investigate emergent memristive dynamics in nanowire (NW) networks by developing a multiterminal characterization setup. IV and pulse characterizations have been performed to characterize the electrical behavior of the network, with great focus on the turn-on phase. Thanks to multielectrode characterization, features like tentative turn-ons and network electrical symmetry have been successfully analyzed. Thanks to the FPGA controlled measurement hardware, a conductivity matrix routine has been implemented, which in a couple of seconds is able to perform measurements of all the bias couple combinations in a 16 electrodes setup. This allowed to have a relatively fast snapshot of the conductive state of the system. The obtained results showed the change in network conductive state before and after electrical stimulation, highlighting the potentiation of the network. For what concerns reservoir computing tasks, one of the key parameters that defines the capability of retaining information of a neuromorphic system is the so called Memory Capacity. In this work, this parameter (together with the forgetting curve of the system) has been experimentally obtained, with comparable results w.r.t. literature simulations. This is a successful outcome that paves the way for the characterization of these networks for reservoir computing purposes. The key highlight of this work is the result obtained with complex experimental reservoir computing tasks such as pattern recognition. The network’s output nodes have been monitored, and their state has been evaluated by a software readout layer. This allowed up to 85% accuracy on diagonal, horizontal, and vertical pattern classification tasks for a 4x4 pixel image. Despite the test set being rather small, the obtained result is outstanding and lays a solid foundation for the application of far more complex tasks such as speech recognition, image classification and time-series prediction.