Artificial visual neuron based on threshold switching memristors

The human visual system encodes optical information perceived by photoreceptors in the retina into neural spikes and then processes them by the visual cortex, with high efficiency and low energy consumption. Inspired by this information processing mode, an universal artificial neuron constructed with a resistor (R s) and a threshold switching memristor can realize rate coding by modulating pulse parameters and the resistance of R s. Owing to the absence of an external parallel capacitor, the artificial neuron has minimized chip area. In addition, an artificial visual neuron is proposed by replacing R s in the artificial neuron with a photo-resistor. The oscillation frequency of the artificial visual neuron depends on the distance between the photo-resistor and light, which is fundamental to acquiring depth perception for precise recognition and learning. A visual perception system with the artificial visual neuron can accurately and conceptually emulate the self-regulation process of the speed control system in a driverless automobile. Therefore, the artificial visual neuron can process efficiently sensory data, reduce or eliminate data transfer and conversion at sensor/processor interfaces, and expand its application in the field of artificial intelligence.


Introduction
The human visual system is composed of the sensory organ (the eyes) and parts of the central nervous system (the retina containing photoreceptor cells, optic nerves, optic tracts, and visual cortex), which gives humans the ability to detect and process visible light. Such a system is a combination of information perception, storage, and computation, which has the advantages of adaptive information processing in real time with high efficiency and low energy consumption [1]. Therefore, the human visual system demonstrates excellent performance in abundant complex tasks, including assessments of depth perception, motion perception, analyses and integrations of visual information, pattern recognition, and more [2].
In contrast to the human visual system, existing visual perception systems use completely different architectures, in which sensors, memories, and computing units are separated from each other [3][4][5]. The information flows as follows: sensors perceive external optical signals and convert them into electrical analog signals with a lot of redundant information; after conversion by an analog-digital converter and amplification by an amplifier, the digital signals are transmitted to remote computing units for further processing [6,7]. With increasingly complex intelligent tasks, such a traditional architecture brings about varied problems such as additional power consumption, long response time, substantial data storage, and data security problem, which are no longer suitable for the Internet of Things [8,9]. Hence building artificial visual perception nervous systems that integrate information perception, storage, and computation is of significant implication for future electronics, in which an artificial visual neuron with the functions of information perception and spike encoding is the prerequisite. Many efforts have been made to realize artificial neurons [10][11][12][13][14][15][16][17]. Recently, threshold switching memristors featuring good scalability and integration density have been demonstrated [11,16,18]. The intrinsic ionic dynamics enables threshold switching memristors to emulate the self-oscillation behaviors of neurons upon applying voltages. Oscillation parameters such as frequency or phase can be used to store the input information of voltage or conductance. Using artificial photoelectric synapses as photoreceptors and spiking encoders based on threshold switching memristors, different researchers constructed several artificial visual neurons [19][20][21]. However, the signal adaptation between the photoelectric synapses and the spiking encoding units needs to be improved. In addition, capacitors occupy a large chip area in neuromorphic hardware systems. Therefore, simplified circuits are necessary.
In addition, threshold switching memristors can also be used to emulate the functions of biological synapses [22][23][24][25]. Sun et al reported Co-Ni layered, double hydroxide-based memristors, which were used to emulate synaptic functions in biology, including short-term plasticity (STP), long-term plasticity (LTP), paired-pulse facilitation (PPF), and spiking-timing-dependent plasticity (STDP) [25]. Wang et al provided a one-selector one-resistor (1S1R) array consisting of selectors and threshold switching memristors, which is an important way to achieve high-density storage and neuromorphic computing [22].
In this work, memristors with Pt/NbO x /TiN-sandwiched structure show excellent threshold switching properties and fast turn-on/off delay time (35/40 ns). Moreover, the threshold switching device's highly stable threshold voltage (V th ) and hold voltage (V hold ) help to improve the distribution of the output frequencies and suppress the degradation of the neural network recognition accuracy. Additionally, an artificial neuron composed of a threshold switching memristor and a resistor in series exhibits outstanding self-oscillation behavior, where the oscillation frequency can be regulated by the applied pulse amplitude and load resistance. More importantly, an artificial visual neuron comprising a photo-resistor and a threshold switching memristor can achieve the self-oscillation behavior as well. To save chip area, no additional capacitor is applied. Compared with photoelectric synapses, photo-resistors have the advantages of simple manufacture, lower cost, and better stability and regulation. The oscillation frequency of the artificial visual neuron depends on the distance between the photo-resistor and light, which is fundamental for applications in the meeting scenarios of driverless automobiles. Finally, a nighttime meeting process is simulated based on the proposed artificial visual neuron.

Results and discussion
The structures of the human eyes and the visual pathway of the human visual system are shown in figure 1(a). The retina in the eyes perceives and converts the external visual information to neural spikes, and the optic nerve transmits the spikes to the visual cortex. Then the visual cortex acts as a computing unit, which integrates and processes the afferent information and transfers it to effectors by efferent nerves to respond to external or internal environments. Figure 1(b) reveals the information processing mode in the retina. The photoreceptors in the retina are mainly responsible for perceiving visual inputs, and the bipolar cells and ganglion cells can, in turn, convert the sensing signals into electrical spikes. Inspired by the information processing mode of the human visual system, an artificial visual neuron consisting of an artificial receptor and a spiking encoder is shown in figure 1(c). By connecting a photo-resistor and a spiking encoder based on a Pt/NbO x /TiN memristor in series, the artificial visual neuron can achieve similar sensing and oscillation behaviors of a biological neuron.
As a key component in the spiking encoder, the Pt/NbO x /TiN memristor is schematically illustrated in the inset of figure 2(a). Figure 2(a) shows the cross-sectional scanning transmission electron microscopy (STEM) image of the device: the NbO x layer is sandwiched by the Pt and TiN electrodes, W is the adhesive layer, and the boundaries between each layer are clear. The spatial mappings of Pt, Nb, O, N, Ti, and W elements are shown in figures 2(e)-(j); there is no reaction between the layers at the interfaces. To further characterize the composition of the NbO x film, a NbO x film was deposited on a silicon wafer for x-ray photoelectron spectroscopy (XPS) analyses. To avoid the effects of surface contamination, the sample was surface etched by 10 nm using Ar + sputtering in the XPS chamber before collecting the XPS information. To confirm the contents of oxygen-related defects in the NbO x film, the fitting results of O1s were analyzed, as shown in figure 2(b). The fitting results show that the Nb-O bonds (purple curve) are located at 530.64 eV, and other non-bridging oxygen bonds (blue curve) are located at 532.50 eV. Figure 2(c) shows the fitting results of N1;, the peak located at 397 eV indicates the existence of N element in the NbO x film. The fitting results with three doublets of the Nb 3d spectrum are shown in figure 2(d) where the lower doublet (purple curve) is clearly identified as oxide Nb 2 O 5 , while the two peaks (green curve) located at 204.26 and 207.04 eV are associated to nitride NbN; the two peaks (yellow curve) located at 205.56 and 208.51 eV correspond to NbO 2 3d5/2 and 3d3/2, respectively, which is consistent with literature reports [26][27][28][29]. Therefore, the reactive sputtered NbO x film contained NbO 2 , Nb 2 O 5 , and NbN phases. It can be explained as follows: during the deposition of the NbO x film in the Ar and N 2 mixture, NbN was formed. After exposure to air for a period of time, sectional NbN reacted with oxygen to form the NbO 2 phase. The unstable NbO 2 phase tended to be oxidized spontaneously to form the more stable Nb 2 O 5 phase in the presence of a high oxygen content of a near surface. As-fabricated devices were initially in a high-resistance state (HRS; ∼10 12 Ω) and an electroforming process was required to initiate threshold switching, as shown in the gray line of figure 3(a); during the electroforming, the NbO x film underwent a spontaneous decomposition into a threshold region and a memory region. The threshold region is dominated by the insulator-metal transition (IMT) in NbO 2 , while the memory region that is in series with the threshold region is composed of oxygen vacancy filaments in Nb 2 O 5 [30]. The DC I-V characteristics of the electroformed device with a compliance current of 2 mA are shown in figure 3(a). When the voltage exceeds the threshold voltage (V th ) of 4.6 V, the NbO 2 in the threshold region undergoes IMT at a high temperature due to Joule heating, which leads to a rapid variation in the conductivity, while it changes back when the voltage is below the hold voltage (V hold ) of 3.3 V [31,32]. To explore the conduction mechanism of the device, the sub-threshold I-V characteristic was fitted using the Poole-Frankel model, and the fitting results indicate that the conduction mechanism of the HRS is mainly dominated by Poole-Frankel conduction, which is consistent with previously reported NbO x [33,34].
The on/off sweeping speed of the threshold switching memristor was detected utilizing the pulse measurement, and the relationship between the turn-on delay time and the pulse voltage is shown in figure 3(b). When the pulse voltage is 4.8 V with a width of 1 µs, the turn-on delay time is 460 ns; when the voltage is gradually increased to 6 V, the turn-on delay time decreases to 35 ns, as shown in figure 3(c). The turn-on delay time gradually tends to saturation with increasing pulse voltage, which is due to the inherent delay of the device. Correspondingly, the threshold switching memristor quickly returns to the HRS with 40 ns turn-off relaxation time under a read voltage of 0.2 V. The fast-speed switching behavior of the device is attributed to the fact that only a short-range atomic arrangement is needed for the transition of NbO 2 . The extremely fast-speed switching behavior of the device is beneficial for applications in artificial neurons and logic circuits.
To evaluate the stability of the devices under pulse stimulation, the threshold switching characteristics were tested by consecutively applying 100 triangle pulses (6 V, 2 µs). Figure 3(d) shows the 1st, 50th, and 100th cycles, which almost coincide. When the pulse voltage increases, the current increases gradually and jumps at the moment of IMT. Specifically, the cumulative distribution probability diagrams of V th and V hold extracted from the threshold switching characteristics were analyzed, as shown in figure 3(e). A coefficient of variation is defined as C V = σ/µ to evaluate the variation, where µ and σ are the mean and standard deviation, respectively [16]. This device exhibits excellent uniformity (C V < 5.2%), as compared with other electrical metallization effect (ECM) devices [35]. The HRS and low resistance state (LRS) retention with spike amplitude of 5 V/3 V and spike width of 1 ms are shown in figure 3(f). The inset in figure 3(f) shows a partial enlargement segment. In 1000 cycles, the fluctuation of the HRS and LRS is small, demonstrating that the device has high stability.
The threshold switching characteristics of the devices make it possible to realize spiking encoding behaviors based on the Pearson-Anson circuit [36], whose configuration is depicted in figure 4(a). The artificial neuron is constructed with a resistor (R s ) and a Pt/NbO x /TiN memristor with an intrinsic parasitic capacitor (C). To make the artificial neuron generate spikes, the R s is selected between the HRS and LRS values. Initially, when a voltage V in is applied, the R s and C form a charging loop (CL) [16] and the C will be   charged due to the voltage dividing effect (HRS > R s ). Once the voltage across the C exceeds V th , the device switches from HRS to LRS, and thus the voltage drop on the device (V out ) suddenly decreases (LRS < R s ) [15], then the C and the device form a discharge loop (DL), and the C discharges. Once the voltage decreases below V hold , the device returns to HRS, and the CL will start to be charged again. Applying a pulse with an amplitude of 6.5 V and a width of 20 µs, as shown in figure 4(b), reversible and stable spikes are persistent until the applied voltage is removed. Additionally, the oscillation behavior of the artificial neuron can be regulated by adjusting the pulse amplitude. The oscillation frequency is defined as the spike firing counts of an artificial neuron per unit time, which varies with the pulse amplitude. As shown in figure 4(c), the oscillation frequency increases first and then decreases with increasing pulse amplitude, demonstrating that the artificial neuron can successfully realize the modulated spike frequency characteristics of a biological neuron [37]. With the increase of the pulse amplitude, the integration time gradually increases, while the relaxation time gradually decreases. When the voltage is less than 6.6 V, the extent of the integration time decrease is greater than the extent of the relaxation time increase, consequently, the oscillation frequency increases. Normally, when the variation of the relaxation time exceeds the variation of the integration time, the oscillation frequency begins to decrease [38]. The oscillation frequency is also affected by the series resistor R s , as shown in figure 4(d). The larger the R s resistance, the smaller the current flowing through the CL, the longer integration time in each oscillation. As a result, the oscillation frequency strongly depends on the circuit parameters of the artificial neuron, such as the load resistance and the input voltage, which is fundamental to the construction of artificial visual perception nervous systems. In the human vision system, the key to clear vision is to focus objects at different distances accurately by adjusting the lens thickness, as illustrated in figure 5(a). To emulate the function of the human visual system, an artificial visual neuron is proposed by replacing the R s in the artificial neuron with a photo-resistor R L (light: R l = 10 kΩ, dark: R L = 2 MΩ), as shown in figure 5(b). A coaxial white light LED is used as the light source. By increasing the distance (L) between the LED and the R L from 14 to 17 cm, the light intensity (E) decreases from 1.2 to 0.81 W M −2 , as shown in figure 5(c), which results in increased R L . Figure 5(d) exhibits the oscillation responses to pulses (7 V, 20 µs) with different L. Subsequently, the oscillation frequency extracted from the oscillation responses reduces with increasing distance, as given in figure 5(e). The distance-dependent oscillation frequency enables the artificial visual neuron to realize depth perception for precise recognition and learning [16]. Interestingly, such an artificial visual neuron can be applied in the meeting control system of a driverless automobile.
To further verify the feasibility of the application, a meeting process in which two driverless automobiles drive toward each other during the nighttime is simulated, as shown in figure 6. Before the two automobiles meet, the initial relative velocity of the vehicles is 5 m s −1 . The initial distance between the two automobiles is 10 m. During the meeting process, the distance (L 1 ) between the two vehicles is gradually shortened. The relationship between the perceived E by the photo-resistor (R L ) and the L 1 follows the formulation of E = I L1 2 [19], where I is the luminous intensity. Therefore, the E gradually increases with decreasing L 1 . Subsequently, the R L gradually decreases. When a constant V in of 7 V is applied to the artificial visual neuron, the real-time output frequency (f i ) of the artificial visual neuron increases according to the relationship f i = 1/(t i − t i−1 ). As a result, the meeting control system safely controls the speed of the vehicle according to the equation a = −K √ f i fmax , where a is the accelerating speed of the vehicle, f max is the maximum output frequency of the artificial visual neuron during the meeting process, and K is a coefficient. When the meeting process ends, the E reduces with increasing L 1 , and the R L increases correspondingly. It can be observed in figure 6 the f i will recover to its initial state, and then the driverless automobile returns to its original speed. Therefore, the vehicles can safely drive in a meeting scenario.

Conclusions
A universal artificial neuron constructed with a series resistor (R s ) and a threshold switching memristor can mimic spike encoding behavior. The threshold switching memristor of Pt/NbO x /TiN expresses typical threshold switching (TS) behavior, and can switch between HRS and LRS in an instant (35 to 40 ns). Meanwhile, the oscillation frequency of the artificial neuron can be modulated by pulse parameters and R s . An artificial visual neuron can be similarly constructed by connecting the threshold switching memristor in series with a photo-resistor. This simple structure can save manufacturing cost and chip area. The artificial visual neuron demonstrates the distance-dependent oscillation frequency, which can be used to acquire depth perception for the precise recognition and learning of an artificial visual perception nervous system. During a meeting process of two driverless automobiles, the meeting control system based on the artificial visual neuron can safely and automatically control the speed of a vehicle according to the light intensity of the oncoming vehicle. Therefore, the simple memristor-based circuits can successfully emulate the functions of the human visual system, further expanding the application of memristors in artificial intelligence. Theoretically, by replacing the photo-resistor with other sensors with appropriate resistance values, such as piezoresistive pressure sensors or temperature sensors, the artificial neuron can also be used for sensing and processing other sensory signals.

Device preparation
The Pt/NbO x /TiN memristive devices were fabricated on silicon wafers with 500 nm thermally grown SiO 2 . First, 21 nm thick TiN bottom electrodes, with a 12 nm W adhesive layer, were deposited on the wafer by DC sputtering. Second, a 56 nm thick functional layer was deposited by sputtering using a Nb target in an Ar and N 2 mixture (Ar:N 2 = 17:1). Finally, 67 nm thick Pt was deposited by DC sputtering as the top electrodes. The three layers were all prepared in consecutive steps, and lift-off was applied in every step.

Structure characterizations
A transmission electron microscope (TEM, FEI Tecnai G2 F30) with energy dispersive X-Ray spectroscopy (EDX) was employed to characterize the microstructure of the device. TEM samples were prepared by means of focused ion beam (Helios G4 UX). For material characterization and electrical measurements, XPS was used to measure the chemical composition of the as-deposited NbO x film.

Electrical property measurements
All the electrical measurements were conducted with a Keithley 4200 SCS connected with a Cascade SUMMIT 11000B semiautomatic probe station.

Data availability statement
All data that support the findings of this study are included within the article (and any supplementary files).