Resistive random access memory based artificial neural network for brain-inspired neuromorphic computing

Abstract

Due to the limitations of size reduction in traditional memory and the von Neumann bottleneck of traditional computing systems, neuromorphic computing based on resistive random access memory (RRAM) has become a promising candidate in artificial intelligence algorithms. RRAM consists of metal two electrode layers sandwiching an intermediate functional layer, the functional layer of RRAM offers exceptional versatility in material selection, which allows RRAM devices to exhibit distinct performance characteristics. The principle of RRAM relies on modulating resistance through controlled formation and rupture of conductive filaments in the functional layer via electrical or other stimuli, which enables both resistive switching and multilevel cell (MLC) storage characteristics. These characteristics enable synaptic plasticity-mimicking biological synapses that update their weights through action potential (AP) transmission. This intrinsic similarity makes RRAM particularly suitable for emulating synaptic functions in neuromorphic systems. Inspired by the brain’s architecture, researchers have engineered memristor arrays using RRAM devices to emulate densely interconnected biological synapses. Advanced implementations include three-dimensional (3D) memristors with multi-layer structures and hybrid systems integrating RRAM with complementary hardware/software platforms. These innovations have facilitated the development of artificial neural networks (ANNs) based on RRAM technology, which are now being deployed for high-efficiency data processing and recognition tasks. This paper begins with RRAM materials, providing a detailed analysis of RRAM’s Resistive switching, ultra-low power consumption, reliability, and MLC storage characteristics. It comprehensively summarizes the synaptic plasticity of RRAM based artificial synapses, including short-term plasticity (STP) and long-term plasticity (LTP), classical Pavlov’s dog experiments, and spike-timing-dependent plasticity (STDP) and spike-rate-dependent plasticity (SRDP). Building on this foundation, this paper introduces RRAM based ANNs and divides them into four different types according to its components and spatial structure, and then deeply discusses the application of artificial neural network (ANN) in data processing and recognition. Finally, challenges and outlooks of RRAM for neuromorphic computing are also deeply discussed.