Asynchronous real-time learning in spiking neural network using 3-terminal resistance random access memory

Abstract

Spiking Neural Networks, inspired by the human brain, are promising as they attempt to solve real-life complex problems, such as pattern recognition, at low energy consumption. Resistance Random Access Memory (RRAM) crossbar array to simulate synaptic weight dynamics, combined with external neuron control circuits, presents a promising approach. The crossbar array of the multilevel resistive memory supports more than two states (LRS and HRS), enhancing RRAM's functionality for analog signals and enabling brain-like processing. However, Reading utilizes low voltage to maintain conductance stability, while writing requires high voltage. Hence, a simultaneous, asynchronous read-write, akin to the brain, remains a significant challenge. Although various solutions exist, a simple, areaefficient solution with low circuit overhead is still challenging. In this paper, a 3-terminal Pr0.7Ca0.3MnO3 (PCMO) RRAM is proposed to enable simultaneous writing and reading, overcoming read-write conflicts of two-terminal RRAM. The typical two terminals of resistive 3T-RRAM are used for writing, and the third terminal is for reading, ensuring real-time asynchronous learning operation. Such an SNN with real-time learning can be advantageous as it reduces circuit overhead and the learning time.