Effect of Asymmetric Nonlinearity Dynamics in RRAMs on Spiking Neural Network Performance

Crossbar-based Resistive Random Access Memory (RRAM) array is a promising candidate for fast and efficient implementation of the vector-matrix multiplication, an essential step in a wide variety of workloads. However, several RRAM devices, demonstrating promising synaptic behaviors, are characterized by nonlinear and asymmetric update dynamics, which is a major obstacle for large-scale deployment in neural networks, especially for online learning tasks. In this work, we first introduce a memristive Spiking Neural Network (SNN) with local learning. Then, we study the effect of this asymmetric and nonlinear behavior on the spiking neural network performance and propose a method to overcome the performance degradation without extra nonlinearity cancellation hardware and read cycles. The performance of the proposed method approaches the baseline performance with 1 ∼ 2% drop in recognition accuracy.