Harnessing the Switching Dynamics of Memristive Synapses in Spiking Neural Networks

Memristive devices, two terminal structures able to change their resistance upon voltage application, have been recently spotlighted as key-enabler of hardware neuromorphic computations, because of their non-volatile state retention, compact size and low programming voltage. Furthermore, their complex dynamics in response to pulse stimulation can be exploited to get access to a continuum of resistance states, which can be exploited in bio-inspired neuromorphic computation.

In this framework, we specifically design materials for memristive devices and optimize their analogue and non-linear dynamics. Further, we investigate the impact of the synaptic update dynamics on the training phase in a biologically-plausible and VLSI-compatible spiking neural network. In particular, through system-level simulations, we show that the non-linear memristive dynamics extends the persistence of memories in the network in comparison to linear synapses with similar resolution.

The work is partially supported by the European project H2020-ICT-2015 NEUral computing aRchitectures in Advanced Monolithic 3D-VLSI nanotechnologies (NEURAM³, grant agreement n. 687299).