Validation of Spiking Neural Networks Using Resistive-Switching Synaptic Device with Spike-Rate-Dependent Plasticity

In this work, we have developed a spiking neural network (SNN) using gradual resistive-switching random-access memory (RRAM) synaptic device. The fabricated RRAM devices demonstrated the characteristics of gradually changing conductance with voltage pulses under both positive and negative polarities, which is suitable for imitating the potentiation and depression functions of a biological synapse by an electron device. Featuring the gradual switching characteristics, spike-rate-dependent plasticity (SRDP) inspired by Bienenstock, Cooper, and Munro (BCM) learning rule was confirmed and modeled for synaptic modification in the SNN. Then, the supervised learning of MNIST patterns was performed on the simulated SNNs, by which it has been validated that the proposed resistive-switching synaptic device and SRDP synaptic modification rule can adjust weights accurately in cooperation without necessitating the conventional calculation-based learning scheme in the artificial neural networks (ANNs), such as error backpropagation.