Emerging memristive neurons for neuromorphic computing and sensing

Abstract

Inspired by the principles of the biological nervous system, neuromorphic engineering has brought a promising alternative approach to intelligence computing with high energy efficiency and low consumption. As pivotal components of neuromorphic system, artificial spiking neurons are powerful information processing units and can achieve highly complex nonlinear computations. By leveraging the switching dynamic characteristics of memristive device, memristive neurons show rich spiking behaviors with simple circuit. This report reviews the memristive neurons and their applications in neuromorphic sensing and computing systems. The switching mechanisms that endow memristive devices with rich dynamics and nonlinearity are highlighted, and subsequently various nonlinear spiking neuron behaviors emulated in these memristive devices are reviewed. Then, recent development is introduced on neuromorphic system with memristive neurons for sensing and computing. Finally, we discuss challenges and outlooks of the memristive neurons toward high-performance neuromorphic hardware systems and provide an insightful perspective for the development of interactive neuromorphic electronic systems.

Keywords: Memristive devices; artificial neurons; neuromorphic computing; neuromorphic sensing; spiking dynamics.

Graphical abstract

Figure 1.

(a) Schematic of a biological nerve circuity including three different neurons: sensory neuron, relay neuron and motor neuron. (b) An action potential with ion exchange processes of Na⁺ and K⁺ channels. (c) Hodgkin-Huxley neuron model. (d) Examples of the biological spiking behaviors, including tonic spiking, tonic bursting, irregular spiking, and spike adaptation.

Figure 2.

Schematic of various switching mechanisms underlying memristive devices for neuron demonstration. (a) Valence change. (b) Electrochemical metallization. (c) Phase change. (d) Mott transition.

Figure 3.

Bioplausible memristive neurons with simple spiking dynamic behaviors. (a) Schematic of artificial neuron based on a phase-change device with a plastic synaptic input array. (b) The conductance changes of a phase-change cell when applying a series of crystallizing pulses. (c) The if dynamics in the PCM neuron. Reprinted with permission from [91]. Copyright 2016 Springer Nature. (d) Schematic diagram of the LIF model implementation with a GeTa₄Se₈-based mott memristor and a resistor; the LIF functions thanks to the accumulation of correlated metallic sites. (e) Experimental neuron spiking firing behavior obtained by applying trains of short pulses with various t_ON and t_OFF. Reprinted with permission from [107]. Copyright 2017 WILEY-VCH. (f) Schematic illustration of the artificial LIF neuron circuit with a SiO_xN_y:Ag-based diffusive memristor. (g) Controlled firing response of LIF neuron under different circuit parameters (C_m and R_a). Reprinted with permission from [83]. Copyright 2018 Springer Nature.

Figure 4.

Biophysical memristive neuron with complex spiking dynamics. (a) HH neuristor circuit diagram and I-V characteristics of NbO_x-based mott memristor. (b) All-or-nothing neuron spiking behavior of the neuristor. (c) Experimental spike outputs of the memristive HH neuron with different circuit parameters. Reprinted with permission from [119]. Copyright 2013 Springer Nature. (d) Implementation of a HH neuron with two VO₂-based mott memristors. (e) Typical voltage-controlled (force V) and current-controlled (force I) I-V characteristics of a VO₂-based mott memristor. (f) Partial schematics of the 23 biological neuron spiking behaviors experimentally demonstrated by memristive HH neuron circuit, including tonic spiking, tonic bursting, phasic spiking, spike frequency adaptation, class 1 excitable, and spike latency. Reprinted with permission from [120]. Copyright 2018 Springer Nature.

Figure 5.

(a) Structure diagram of third-order NbO₂-based mott memristor. (b) The equivalent circuit model of the third-order nanocircuit element. (c) Quasistatic I-V behavior of third-order NbO₂-based device. (d) Measured various neuron spiking dynamic behaviors of the third-order NbO₂-based device with different applied external voltages. Reprinted with permission from [102]. Copyright 2020 Springer Nature.

Figure 6.

Comparisons of ANN and SNN. (a) Illustration of non-spiking ANN, where neuron is determined by the activation function to process numerical value input. (b) Illustration of SNN, where neuron is determined by the biological spiking dynamics to process the event-based spike input.

Figure 7.

Hardware implementation of SNN computing based on memristive spiking neurons. (a) Optical micrograph of an integrated full memristive neural network consisting of an 8 × 8 ITIR memristive synapse crossbar and eight memristive spiking neurons. (b) Fully integrated memristive neural network for pattern classification of four letters ‘UMAS’. Reprinted with permission from [83]. Copyright 2018 Springer Nature.

Figure 8.

Hardware implementation of SNN computing based on memristive dendritic neurons. (a) Photograph of the neural network with memristive soma, dendrites, and synapse crossbar. (b) Dendritic neural network computing for image processing; comparisons of the neuron firing rates with and without artificial dendrites. (c) Comparison of the power consumption (left) and recognition accuracies (right) of this hardware system with/without dendrites. Reprinted with permission from [135]. Copyright 2020 Springer Nature. (d) Optical image of the designed dendritic neural network hardware platform on a printed circuit board. (e) Left: accuracy comparison for human motions recognition with/without dendritic functions; right: power consumption comparison for human motions recognition with dendritic functions running on GPU and memristors-based platform. Reprinted with permission from [136]. Copyright 2022 WILEY-VCH.

Figure 9.

Artificial haptic sensory system based on memristive neurons. (a) Schematic of a biological haptic sensory system. (b) Schematic illustration of the power-free artificial spiking mechanoreceptor system based on memristive spiking neuron. Here, the tactile sensory and voltage signal are generated by the piezoelectric device. (c, d) The output signals of the artificial spiking sensory system under different pressure intensities. Reprinted with permission from [29]. Copyright 2020 Springer Nature. (e) Schematic illustration and circuit diagram of the artificial mechanical sensory system composed of a pressure sensor (left) and a NbO_x-based mott memristor (right). (f) Pulse coupled neural network for tactile information sensation enhancement based on the artificial mechanoreceptor. (g) Bio-inspired tactile integration can be successfully implemented by two parallel pressure sensors (P_α and P_β) and a series memristor. (h) Electrical spiking frequency enhanced when external stimuli were applied on both sensors. Reprinted with permission from [139]. Copyright 2021 American Chemical Society.

Figure 10.

Artificial visual sensory system based on memristive neurons. (a) Schematic of the human visual sensory system. (b) The circuit scheme for the visual sensory system based on photoelectric memristive spiking neuron. (c) The firing rate of the left eye and the right eye at a certain distance, respectively. (d) A spiking visual neural network for image recognition based on the artificial binocular visual system. Reprinted with permission from [147]. Copyright 2022 WILEY-VCH. (e) Schematic diagram of the anatomical organization of the visual system with the LGMD neuron. (f) Photograph of the device-level the hemispherical shaped biomimetic compound eye based on flexible Ag/FLBP-CsPbBr₃/ITO-based memristor crossbar, where FLBP stands for few-layer black phosphorous nanosheets and ITO is indium tin oxide. (g) Excitatory and inhibitory response of the memristor to a looming light stimulus with simultaneously applied electric pulses. (h) The artificial LGMD neuron fire pulse number vs. light power. (i) Schematic diagram illustrating the robot car’s decision-making with optical signal processing ability based on memristive spiking neuron. Reprinted with permission from [148]. Copyright 2021 Springer Nature.

Figure 11.

Artificial multisensory system based on memristive neurons. (a) Illustration of human multisensory biological system. (b) Schematic of spike-based artificial neuromorphic sensory system. The calibratable artificial sensory neuron combined with various sensors (pressure sensor, temperature sensor, light sensor, and curvature sensor). (c) Response of artificial spiking sensory neuron to different stimuli (pressure, light intensity, and temperature). Reprinted with permission from [153]. Copyright 2022 Springer Nature. (d) Schematic of the artificial multisensory neuromorphic computing system consisted of a multimode-fused spiking neuron (MFSN) array and an SNN classifier. The MFSN cell is composed of a NbO_x-based mott memristor and a pressure sensor. Reprinted with permission from [137]. Copyright 2022 WILEY-VCH.