Brain-inspired computing with resistive switching memory (RRAM): Devices, synapses and neural networks

• Published in: Microelectronic engineering Vol. 190; pp. 44 - 53
• Main Author: Ielmini, Daniele
• Format: Journal Article
• Language: English
• Published: Amsterdam Elsevier B.V 15.04.2018; Elsevier BV
• Subjects: Brain; Cognition; Computation; Computer memory; Deep learning; Electronic devices; Human performance; Memory; Memory devices; Memristor; Multilevel storage; Nanotechnology devices; Neural networks; Neuromorphic engineering; Pattern recognition; Plastic properties; Power consumption; Resistive switching memory (RRAM); Scalability; Spike-timing dependent plasticity (STDP); Switching theory; Synapses; Multilevel storage; Deep learning; Spike-timing dependent plasticity (STDP); Memristor; Neuromorphic engineering; Resistive switching memory (RRAM)
• ISSN: 0167-9317; 1873-5568
• DOI: 10.1016/j.mee.2018.01.009

The human brain can perform advanced computing tasks, such as learning, recognition, and cognition, with extremely low power consumption and low frequency of neuronal spiking. This is attributed to the highly-parallel and the event-driven scheme of computation, where energy is used only when and where it is needed for processing the information. To mimic the human brain, the fundamental challenges are the replication of the time-dependent plasticity of synapses and the achievement of the high connectivity in biological neuron networks, where the ratio between synapses and neurons is around 104. This combination of high computing capability and density scalability can be obtained with the nanodevice technology, notably by resistive-switching memory (RRAM) devices. In this work, the recent advances in RRAM device technology for memory and synaptic applications are reviewed. First, RRAM devices with improved window and reliability thanks to SiOx dielectric layer are discussed. Then, the application of RRAM in neuromorphic computing are addressed, presenting hybrid synapses capable of spike-timing dependent plasticity (STDP). Brain-inspired hardware featuring learning and recognition of input patterns are finally presented. [Display omitted] •Synapses based on resistive switching memory (RRAM) allow high connectivity and computational efficiency•SiO2 RRAM devices feature high resistance window and multilevel capability which is needed for neuromorphic applications•Unsupervised learning is enabled by spike timing dependent plasticity (STDP) via set and reset processes in RRAM synapses•STDP synapses can be used broadly in feedforward and recurrent networks capable of unsupervised learning