Memory capacity of neuromorphic self-oganizing nanowire networks

Physical Reservoir Computing (PRC) is one of the most promising new computation paradigms oriented to overcome the limitations of the traditional von Neumann architecture. Given their ease of production and low energy consumption, neuromorphic self-organized networks of nanowires with memristive junctions are currently being explored as an interesting physical reservoir for the implementation of PRC paradigm. The recent demonstration of the computing capabilities of these nanowire networks, including recognition of spatio-temporal patterns and time-series prediction, has grown the interest in finding a way to assess the information processing capability of the networks and its dependences on system physical parameters. A possible way to do this is to use the Memory Capacity (MC) benchmark test, already used in the framework of reservoir computing to evaluate the computing capability of the reservoirs. In this work, by means of simulations based on a random graph model of the nanowire networks, it is first discussed a way to properly measure MC, and then the dependence of MC on several system parameters such as network density, network topology, input signal amplitude, and number of reading electrodes is investigated. The aim of these simulations and measurements is to provide a deeper insight of the relations that exists between system parameters and system performances in order to support a more complex experimental activity towards the implementation of PRC paradigm in neuromorphic self-organized nanowire networks.