Low-dimensional optoelectronic synaptic devices for neuromorphic vision sensors

Neuromorphic systems represent a promising avenue for the development of the next generation of artificial intelligence hardware. Machine vision, one of the cores in artificial intelligence, requires system-level support with low power consumption, low latency, and parallel computing. Neuromorphic vision sensors provide an efficient solution for machine vision by simulating the structure and function of the biological retina. Optoelectronic synapses, which use light as the main means to achieve the dual functions of photosensitivity and synapse, are the basic units of the neuromorphic vision sensor. Therefore, it is necessary to develop various optoelectronic synaptic devices to expand the application scenarios of neuromorphic vision systems. This review compares the structure and function for both biological and artificial retina systems, and introduces various optoelectronic synaptic devices based on low-dimensional materials and working mechanisms. In addition, advanced applications of optoelectronic synapses as neuromorphic vision sensors are comprehensively summarized. Finally, the challenges and prospects in this field are briefly discussed.


Introduction
As the development of artificial intelligence and autonomous driving continue to advance, there is a growing need to address issues related to unstructured data and real-time * Author to whom any correspondence should be addressed.
Original content from this work may be used under the terms of the Creative Commons Attribution 4.0 licence. Any further distribution of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI. processing [1][2][3]. Conventional computing architectures based on the von Neumann model are limited by the physical separation of memory and central processing units, making them increasingly inadequate in terms of efficiency and energy consumption for new application scenarios [3][4][5][6]. In contrast, the human brain represents a sophisticated 'intelligent computer', in which storage and computing are integrated within the same unit [7][8][9][10]. The distributed, parallel, low power consumption, and high error tolerance characteristics of neural networks composed of hundreds of millions of neurons and synapses enable us to perform various intelligent tasks, including language processing and image recognition [11][12][13][14]. In particular, the retina and visual cortex of the human brain, which is responsible for the majority of human information processing, are highly compact and energy-efficient [15][16][17]. In light of these advantages, brain-inspired artificial neural networks (ANNs), as proposed by Mead in 1990, are considered to be one of the most promising technologies for overcoming the von Neumann bottleneck [18].
The human brain comprises of a complex neural network composed of 10 11 neurons and 10 15 synapses [19,20]. Synapses, which are the connections between neurons, serve as the basic units of computation and learning. Therefore, simulating the biological functions of synapses is the key to developing neuromorphic computing. In recent years, various structures and principles of electronic synaptic devices have been proposed, such as memristors [21][22][23][24][25][26], phase change memory [27][28][29][30][31], resistive-switching memory [32,33], and ferroelectric transistors [34][35][36]. Despite significant progress, electronic synapses remain limited in terms of operating speed, bandwidth density, and interconnectivity, given their typically pure electronic inputs and outputs. By contrast, optoelectronic synapses simulate synaptic function through optical or electrical stimulation, and offer distinct advantages such as non-contact operation, low crosstalk, large bandwidth, and low power consumption [37][38][39]. Furthermore, the exceptional optical response of optoelectronic synapses enables them to simultaneously combine visual sensing, information storage, and computing, and thereby simulate the visual perception and information processing functions of the human brain. This feature is especially significant for constructing neuromorphic vision sensors. Therefore, in recent years, a range of optoelectronic synaptic devices simulating functions of the human visual systems have been proposed, including color recognition [40][41][42][43][44][45][46], angle recognition [47][48][49][50], contrast enhancement [51][52][53][54][55], motion detection [56][57][58][59][60][61], and visual adaptation [62][63][64][65][66][67], which will be discussed in detail in section 5. In this review, we discuss the progress of low-dimensional optoelectronic synaptic devices in neuromorphic vision sensors. First, we introduce the structure and principles of the retina, which shows the progressiveness of the biological vision systems. Then, the basic functions of synaptic plasticity are reviewed in detail. Secondly, we briefly discuss the working mechanisms and low-dimensional material systems of several optoelectronic synapses. Finally, we provide a comprehensive overview of advanced neuromorphic vision sensor applications and propose the current challenges and prospects for the development of neuromorphic vision sensors in the future.

Biological retina and synaptic plasticity
The human visual system is highly efficient, with hundreds of millions of photosensitive neurons on the retina responsible for light sensing and signal pre-processing. The neurons are capable of detecting and filtering out redundant information and transmitting key information to the visual cortex of the brain for deeper processing [68,69]. The goal of neuromorphic vision sensors is to emulate the human visual system by using photosensitive devices to integrate perception, storage, and computing. At the heart of neuromorphic vision sensors are optoelectronic synapses, which are capable of achieving basic visual functions, such as contrast enhancement and image recognition by tuning the conductance through electrical or optical spikes, thus simulating synaptic plasticity [70,71]. In this section, we focus on introducing the structure and function of the retina, as well as the plasticity of optoelectronic synaptic devices.

Retinal structure and function
Visual information accounts for about 80% of human access to external information [72]. The retina, a key component of the visual system, is responsible for transforming optical signals that carry visual information into electrical signals that are then pre-processed by the retina's complex network structure and transmitted to the visual cortex for further processing [73,74] (figure 1(a)). Comprised of a pigment epithelial layer and a retinal sensory layer, the retina is structurally a soft and transparent film with thin edges and a thick middle [75]. The pigment epithelial layer supports and nourishes photoreceptor cells, while also performing important functions such as shading heat dissipation, regeneration, and repair. The retinal sensory layer, on the other hand, consists of five main types of cells: cone and rod cells, bipolar cells, horizontal cells, amacrine cells, and retina ganglion cells [76,77].
As shown in figure 1(a), these cells are arranged in a specific sequence, with the outer neural layer (ONL) containing photosensitive cells such as rod cells and cone cells. The ONL captures incoming light and converts it into electrical signals, which are then transmitted to the outer plexiform layer, which contains the axons of photoreceptor cells and dendrites of bipolar cells, forming the first synaptic layer. Signals are further processed by horizontal cells and then transmitted to bipolar cells, which form parallel information pathways representing different conversions of photoreceptor signals, providing highly pre-treated excitatory input for the internal retina. The inner neural layer contains the cell bodies of bipolar cells, horizontal cells, and amacrine cells. The inner plexiform layer contains the axons of bipolar cells, amacrine cells, and dendrites of retinal ganglion cells, with amacrine cells providing major inhibitory or neuroregulatory inputs for bipolar cells and retinal ganglion cells. Finally, the ganglion cells layer integrates the inputs of different sets of bipolar cells and amacrine cells, and encodes the results into a series of spikes, which are sent to the higher visual center through their axons to form the optic nerve.
This intricate layered structure enables the retina to encode optical signals carrying visual information into electrical signals via photoreceptor cells, while the detailed local interactions between bipolar cells, amacrine cells, and retinal ganglion cells within the retina, support the retina's fundamental visual feature extraction ability.

Synaptic plasticity of optoelectronic neuromorphic devices
Neuromorphic vision sensors realize the functions of perception, memory, and computing by simulating the information transmission between neurons in the retina [15,52,78] ( figure 1(b)). In the biological nervous system, neurons consist of dendrites, nerve cell bodies, and axons. The information can be transmitted between adjacent neurons through synapses, and the strength of the connections between neurons is referred to as the synaptic plasticity, which can change the synaptic weight [79,80]. Neural networks achieve functions such as learning, forgetting, image recognition, and logic by adjusting the synaptic weight value [70,[81][82][83][84], making synaptic plasticity the primary criterion for evaluating neuromorphic devices. Synaptic plasticity has many forms, including short-term plasticity (STP), long-term plasticity (LTP), spike-timing-dependent plasticity (STDP), and spiking-ratedependent plasticity (SRDP). STP and LTP depend on the length of the memory time. STP, also known as dynamic synapses, refers to the change of synaptic efficiency over time, which reflects the history of presynaptic activity [85].
Paired-pulse facilitation (PPF) and paired-pulse depression (PPD) are the most well-known functions of STP, in which the response of the synaptic device to the second stimulus can increase or decrease compared to the response to the first stimulus, resulting in PPF or PPD. The effect of STP on synapses is temporary, typically lasting between hundreds and thousands of milliseconds, and its impact on synaptic efficiency will soon return to the ground state level without continuous presynaptic activity. However, STP can gradually transform into LTP after several iterations of external information training. Generally, optoelectronic synaptic devices use optical or electrical spikes as stimuli to change STP to LTP by changing frequency, quantity, or amplitude.
Although STP and LTP are the basic characteristics of biological synapses, they cannot accurately describe the learning rules of brain neural networks. To understand the interaction between neurons and explore the interconnection between multiple synapses, STDP and SRDP have been proposed as advanced learning rules in the human brain, representing the basic mechanism of memory and learning [86,87]. STDP describes the time sequence and time interval dependence between the two spikes of presynaptic and postsynaptic neurons. Four forms of STDP exist according to the time sequence and time interval. When presynaptic neurons are stimulated earlier than postsynaptic neurons, the postsynaptic current is enhanced. When presynaptic neurons are stimulated later than postsynaptic neurons, the postsynaptic current is inhibited. Meanwhile, changing the time interval between two stimuli also affects the size of the postsynaptic current. STDP can be considered as the competitive spike formula of Hebbian learning rules because it introduces competition between different synapses to control the time of postsynaptic action potential. In contrast, SRDP changes the sign and size of synaptic plasticity by changing the emission rate of presynaptic spikes. High-frequency emissivity (20-100 Hz) leads to potentiation, while low-frequency emissivity (1-5 Hz) leads to depression. SRDP reveals activity-dependent synaptic plasticity, indicating that low (high) postsynaptic activity reduces (enhances) synaptic efficiency.

Mechanisms for optoelectronic synaptic devices
Optoelectronic synaptic devices differ from traditional photoelectric devices in that they can maintain a certain conductivity memory even after the removal of light excitation, in addition to exhibiting rapid light response that changes electrical conductivity [88,89]. This memory performance is a key indicator for achieving synaptic plasticity, which is influenced by light excitation frequency, pulse width, and excitation times. Optoelectronic synaptic devices contain multiple mechanisms, such as photo-stimulated phase changes [90], ion migration [91], floating gate-based [92,93], trapping and detrapping [94,95], and so on. It is necessary to carefully design the material ′ s composition and control the transmission and recombination of photogenerated carriers to achieve biological functions. In this section, we focus on the trapping and detrapping process of optoelectronic synapses, including defect engineering and band engineering, as important approaches to improve the performance of these devices.

Defect engineering
Semiconductor defects include surface defects, interface defects, and body defects, which play an effective modulation role in carrier transport [96][97][98][99]. Defect engineering is considered as a simple method to generate new functionalities in material science, particularly for simulating biological synaptic functions in optoelectronic synaptic devices [100]. Under light excitation, electrons in the valence band of semiconductor materials are promoted to the conduction band, forming electron-hole pairs. External electric fields can collect some of these electron-hole pairs at the electrode, while others may be trapped by defect states. Although trapped electrons can be de-trapped due to the thermal motion, this process requires a relaxation time. As a result, after turning off light excitation, the electrical conductivity of the device cannot return to its original level, thereby generating excitatory or inhibitory postsynaptic currents. The number of defect states, energy level depth, and light excitation conditions can be appropriately designed to further adjust the synaptic weight. For instance, nanowire (NW) materials have large surface area-to-volume ratios and abundant surface states [101][102][103]. Under light excitation, many photo-generated carriers can be trapped by the surface state, resulting in an increased carrier lifetime and the formation of the persistent photoconductivity (PPC) effect. Black phosphorus (BP) materials spontaneously produce numerous defect states [104,105]. Different UV wavelengths can stimulate the generation of positive and negative photoconductivity, simulating excitatory and inhibitory action potentials in synapses, respectively. Additionally, most two-dimensional (2D) materials are constructed on SiO 2 /Si substrates, which possess abundant defect states on their surfaces [106][107][108][109]. Under optical or electrical excitation, these defect states can capture and remove a large number of electrons, thereby adjusting synaptic weight.

Band engineering
Band engineering is a method used to control the process of trapping and de-trapping photo-generated carriers by designing and modulating band alignment. The creation of potential barriers and internal electric fields at heterojunctions and Schottky junctions can promote or inhibit the separation and recombination of photo-generated carriers. With the aid of band engineering, the capture and recombination of photogenerated carriers can be finely tuned, allowing for precise adjustment of the weight of optoelectronic synapses.

Type I and II heterojunction.
Type I and II heterojunctions are widely used in band engineering to separate photogenerated carriers in optoelectronic devices [110][111][112]. The use of heterojunctions to adjust the optical and electrical properties of materials has become a popular research field. The heterojunction band alignment principle has been applied to many optoelectronic synaptic devices. As shown in figure 2(a), Xie et al reported the type I heterojunction optoelectronic synaptic device composed of CsPbBr 3 and MoS 2 [67]. Under a positive gate voltage, most of the electrons were driven from CsPbBr 3 to MoS 2 due to the bending of the energy band, leading to a significant increase in conductivity in MoS 2 . Subsequently, time-dependent electron trapping occurred in the space charge region, resulting in a final reduction in electron concentration. Upon removal of the gate stimulation, rapid electron de-trapping occurred, restoring the device performance to its original state. Type II heterojunctions are considered to be one of the most suitable device types for optoelectronic synapses. Wang et al reported a kind of MoS 2 /perylene-3,4,9,10-tetracarboxylic dianhydride (PTCDA) hybrid II heterojunction optoelectronic synaptic device [113]. Through careful design of the energy band arrangement, the laser pulse produced a large number of unbalanced electrons transferred from PTCDA to MoS 2 , resulting in a sharp increase in PSC under the light modulation mode. After the laser pulse was removed, the electrons gradually returned to PTCDA, leading the corresponding PSC to the initial state, constituting the excitatory post-synaptic current (EPSC) behavior. The photogating effect and charge transfer under laser pulse are also shown in figure 2(b).

Metal-semiconductor contact barrier.
Photo-excited electron-hole pairs can significantly impact the conductivity of devices by altering carrier concentration at the semiconductor and interface, which can in turn affect the contact barrier [115][116][117][118][119]. Among the mechanisms of the optical modulation barrier, the metal-semiconductor interface contact barrier is the most common. Recently, Gao et al demonstrated the construction of Schottky optoelectronic synaptic devices using ITO and Nb:SrTiO 3 [114]. By employing a mixed pulse spike composed of electricity and light, the height and width of the Schottky barrier can be modulated, allowing for dynamic adjustment of synaptic conductance via electron trapping and de-trapping processes, which was assisted by the abundant oxygen defects in Nb:SrTiO 3 . The working mechanism is illustrated in figure 2(c). The energy band level mismatch between ITO and Nb:SrTiO 3 resulted in the formation of a contact potential barrier at the interface. The barrier impeded electron transmission, causing the ITO/Nb:SrTiO 3 heterojunction to initially exhibit a high resistance state. This, in turn, decreased the height/width of the contact barrier, thereby increasing the probability of electron tunneling and inducing a switch in conductivity to the low resistance state.

0D materials
The proliferation of neuromorphic sensing and computing has spurred the development of high-performance brain-inspired neuromorphic hardware. Quantum dot (QD) materials offer an attractive option for such applications in semiconductor technology. QDs are typically sized close to, or smaller than the exciton Bohr radius, which exhibits significant quantum confinement effects and photoluminescence efficiency [161,162]. By controlling their size, shape, and surface ligand, devices with desired electrical and optical properties can be tailored. The concentration and dynamics of excitons in QDs can be adjusted by input optical pulses, which can enhance device performance, including reducing power consumption, increasing conductivity, accelerating switching speed, and enriching device operability. Furthermore, the production cost of QD materials is low, which can meet the requirements of largearea and solution-based manufacturing [163]. In particular, the 0D materials can be treated in solution while maintaining the excellent optoelectronic properties and structural stability of crystalline inorganic materials, providing new opportunities for neuromorphic devices.
The silicon nanocrystal (NC) is an important nanostructure with broad absorption efficiency from ultraviolet to nearinfrared under heavy doping, demonstrating significant optical properties [127,128,164]. Tan et al reported a B-doped silicon nanocrystalline optoelectronic synaptic transistor, which exhibits various typical synaptic functions under the excitation of multi-wavelength light pulses from ultraviolet to nearinfrared [127]. The plasticity of the synapse arises from the dynamic capturing and releasing of photogenerated carriers at defects, such as dangling bonds, on the surface of NCs. Moreover, the research group successfully demonstrated the aversion to learning behavior and handwritten character recognition using optoelectronic co-stimulation, demonstrating the potential of nanocrystalline materials in neuromorphic perception and computing [128]. Among many semiconductor QDs, inorganic perovskite QDs possess excellent spectral absorption efficiency, relatively low carrier defect density, long carrier lifetime, and narrow exciton binding energy, making them promising candidates for optoelectronic synaptic devices. Photonic flash devices based on all-inorganic CsPbBr 3 perovskite QDs have been shown to simulate typical synaptic functions, such as STP, LTP, and STDP [126]. Furthermore, due to the excellent broadband response of CsPbBr 3 , synaptic weight exhibited multi-wavelength modulation characteristics. Real-time recognition/decoding of visual information was closely related to the PPF behavior of post-synaptic neurons. Han et al reported a CsPbBr 3 /h-BN/Gr light-stimulated synaptic transistor that blocked photogenerated holes due to the existence of ultra-thin h-BN, resulting in a large PPF  [125]. Based on the pixel array, the functions of learning and forgetting and handwritten character recognition have been demonstrated. These results highlight the significant potential of perovskite QDs in optoelectronic synapses and artificial visual perception systems.

1D materials
One-dimensional materials, such as NWs and nanorods, exhibit unique characteristics including high aspect ratio, large surface area volume ratio, and sub-wavelength size effects, which enable carriers to move freely on a single nanometer scale. The 1D structure results in high carrier mobility, high photoelectric conversion efficiency, and adjustable optical absorption coefficients [165][166][167][168], allowing for a wide spectrum of detection capabilities from ultraviolet to infrared waves. The simple and low-cost preparation process of semiconductor NWs, such as the vapor-liquid-solid (VLS), enables the transfer of NWs to any substrate or material to form flexible devices or heterostructures. In recent years, a series of 1D materials represented by carbon nanotube (CNT) [129,130], SnO 2 NWs [131,154], ZnO NWs [98,[155][156][157], TiO 2 NWs [158], and InAs NWs [132] have been reported to be used in optoelectronic synaptic and neuromorphic devices. CNT, the most widely studied 1D material, has excellent physical and chemical stability, high electron and hole mobility, and a wide spectral detection rate. Single-walled CNTs (SWCNTs), as excellent ink materials, have been utilized in the fabrication of high-quality SWCNT TFTs for the display, sensing, and wearable devices. Optoelectronic neuromorphic devices based on printed SWCNT TFTs demonstrated rapid response to 520-1310 nm light pulses and good PPC effect, leading to the realization of important synaptic functions including learning, memory, and signal filtering [129]. SnO 2 NWs were attracted as the optoelectronic devices due to the wide band gap (3.5-4.0 eV), high aspect ratio, and large amount of oxygen vacancy density that created surface states and internal defects. These features led to trapped holes, which increased the electron density and the conductivity. The PPC effect of SnO 2 NWs has been used to simulate basic biological synaptic function [131]. In addition, ZnO NWs have excellent optical properties, low cost, and ease of manufacture, making them suitable for artificial optoelectronic synaptic devices. As the non-centrally symmetric crystal structures, ZnO NWs exhibited significant piezoelectric effects under the strain. An artificial synaptic device based on ZnO NWs piezoelectric photoelectric effect simulated the important synaptic functions, such as PPF, and STP/LTP under the UV light pulse [156]. Additionally, the transport of photogenerated carriers can be controlled via piezoelectric effects to adjust the synapse weight, resulting in the great application potential of 1D materials in neuromorphic photonics.

2D materials
The unique atomic-level ultra-thin properties of 2D materials, composed of covalent bonds and weak van der Waals forces, results in confined electron behavior within the 2D plane and energy dependence on the number of layers. For instance, monolayer MoS 2 exhibits a direct band gap, while bilayer MoS 2 has an indirect band gap, making these materials attractive for their excellent electrical and optical properties, including wide spectral response, atomic scalability, fast optical response, high optical quantum efficiency, and electrical tunability [169]. Moreover, the reasonable design of device structures has enabled the observation of a photoinduced PPC effect, which is critical for the development of optoelectronic synapses based on 2D materials. Simple preparation processes, such as chemical vapor deposition, have produced impressive sizes of 2D materials, and further integration with complementary metal-oxide-semiconductor (CMOS) processing technology can maximize their potential in high-performance and reliable optoelectronic devices, providing a realistic technical basis for the development of large-scale 2D artificial vision chips.
As a member of the graphene-like family, BP exhibits high mobility, adjustable band gap, and wide spectral response, making it an ideal 2D optoelectronic synaptic material. However, BP spontaneously oxidized in the natural environment, results in abundant defect states that can alter its intrinsic photoelectric characteristics. Nevertheless, the rational use of spontaneous defects can create unique optoelectronic functions, such as wavelength-selective positive and negative photoconductivity based on BP defect engineering [105,159,160]. This adjustable characteristic of conductivity polarity mimicked the function of bipolar cells in the human retinal system. By exploiting these optically adjustable positive and negative photoconductivity characteristics, it is possible to simulate excitatory and inhibitory action potentials, as well as various synaptic functions, such as the transition from short-term to long-term memory and PPF. Unfortunately, controlling the natural oxidation of BP is challenging, and device-level BP can only be prepared by micromechanical exfoliation, which limits its application at the wafer-scale level. Transition metal chalcogenides, such as MoS 2 , were regarded as ideal optoelectronic synaptic materials due to their wafer-scale preparation, CMOS process compatibility, and flexible hetero-integration [170,171]. MoS 2 was utilized as an effective light absorption layer and charge capture layer in optoelectronic synaptic devices, and its flexibility made it suitable for wearable devices and artificial retina bionics. Several optoelectronic synaptic devices based on MoS 2 materials have been reported, including multi-grid liquid ion gate MoS 2 [138], in-enhanced electron implantation MoS 2 [140], artificial grain boundary engineering MoS 2 [141], Au nanoparticles modified MoS 2 [143] and flexible MoS 2 [137]. These devices have demonstrated excellent performance in synaptic plasticity, pulse power consumption, and flexible sensing, with good potential for building the human visual system. However, the atomic-level ultra-thin characteristics of 2D materials make it challenging to introduce dopants into them effectively, limiting their precise modulation of electrical properties and power consumption.

Heterostructures
Heterojunctions represent a highly versatile strategy to fully exploit the advantages of different materials and integrate them to produce unique physical properties, including high electron mobility, excellent optical absorption, superior quantum efficiency, and extended exciton lifetime. Therefore, heterojunctions have emerged as a highly attractive option in the field of artificial synapses and neuromorphic vision. At present, two main types of heterojunction optoelectronic synaptic devices are widely used, namely 0D/2D heterojunctions and 2D/2D heterojunctions. In the 0D/2D system, the 0D materials served as a photosensitizer due to their excellent optical properties resulting from the quantum confinement effect, while the 2D materials acted as a charge transfer layer thanks to the non-hanging bond and weak interlayer interaction. Under external light stimulation, the 0D/2D heterojunction delivered exceptional optical response and high optical carrier separation efficiency [66,144]. On the other hand, in the 2D/2D system, the van der Waals effect weakened the lattice matching constraint, allowing for the stacking of any arbitrary materials and the consequent enrichment of device operation and physical characteristics. The absence of defects at the interface and the complementary optical absorption properties rendered this system highly electrically and optically efficient. The low energy consumption and enhanced stability of device operation can be achieved by the thin geometry and chemical inertness of the constituent atoms [172]. Additionally, the gate-adjustable interface facilitates charge transfer and transmission, thereby enabling unique optical response characteristics and the PPC effect.
Metal halide perovskite has emerged as a highly attractive photosensitizer for 0D/2D optical synapses due to its superior light absorption, small exciton binding energy, and long carrier lifetime. Hong et al reported an optoelectronic synaptic device composed of 0D CsPb(Br 1−x I x ) 3 and 2D MoS 2 [66]. As a channel for light sensing, MoS 2 exhibited weak light absorption owing to its atomic-level thickness. The incorporation of CsPb(Br 1−x I x ) 3 enhanced the light absorption of the structure. The hybrid structure with photo-induced phase separation led to a time-dependent change in photocurrent in the phototransistor, further simulating the adaptive function in neuromorphic vision sensors. Recently, Liang et al reported a photo-synaptic transistor based on InP/ZnSe QDs and SnO 2 heterojunction [146]. The reasonable band arrangement in the QDs/SnO 2 structure allowed for the effective separation of photogenerated electrons and holes at the heterojunction interface. Meanwhile, the distribution of oxygen vacancy defects in the SnO 2 film delayed the electron recombination and enhanced the PPF behavior. QD/SnO 2 optoelectronic synaptic devices also displayed a variety of typical synaptic functions, such as STP and LTP. Importantly, all components of the QDs/SnO 2 devices were produced by a solvent-based inkjet printing process, offering a cost-effective and highyield strategy for constructing neuromorphic devices. The band arrangement of 2D/2D heterojunctions imparted unique photoelectric characteristics. Recently, neuromorphic devices based on gate-tunable positive and negative photoconductivity have been proven [55]. In the WSe 2 /h-BN/Al 2 O 3 structure, positive and negative photoconductivity was achieved by electrostatic doping through grid regulation to simulate the characteristics of human bipolar cells. The visual sensor was further used as a convolution neural network for image recognition to enable the dynamic detection of objects. In addition, the introduction of ferroelectrics into the synaptic structure provided new degrees of freedom to the device's function. Recently, ferroelectric 2D materials α-In 2 Se 3 and GaSe heterojunction synapses have been reported to simulate the human visual system [147]. α-In 2 Se 3 exhibited higher conductivity and a narrower band gap than oxide ferroelectrics and displayed strong ferroelectricity at room temperature. The α-In 2 Se 3 /GaSe heterojunction device can simulate the typical PPF, LTP, and LTD synaptic functions. Under light stimulation, wavelength selectivity was achieved, demonstrating the color recognition ability of the artificial retina. Moreover, associative learning, logic operation, and image recognition functions have been realized, underscoring the strong potential of 2D/2D heterojunctions in neuromorphic vision sensors.

Advanced applications in neuromorphic vision sensors
Neuromorphic vision sensors aim to simulate the human retina and its visual system, with the functions of sensing, storing, and preprocessing light signals. Compared with conventional digital image systems, neuromorphic vision sensors had high parallelism and low power consumption while filtering most redundant information. Visual signals were carried by light, which contains wavelength (color) and spatial direction information. Upon entering the retina, the optical signal underwent further preprocessing to achieve noise reduction and contrast enhancement. Furthermore, time-dependent optical signal inputs enabled spatial movement detection of objects while maintaining a certain visual threshold range, allowing for selfadaptive functionality in strong or weak light environments to achieve accurate perception. In this section, we summarize the advanced applications of neuromorphic vision sensors from five perspectives: color recognition, angle recognition, contrast enhancement, motion detection, and visual adaptation, showcasing the considerable potential and development of neuromorphic vision sensors.

Color recognition
The human visual system can recognize the wavelength information of light to form color recognition. Similarly, optoelectronic synaptic devices can achieve wavelength selectivity by having different postsynaptic currents due to the different optical absorptivity and external quantum efficiency of materials for multiwavelength light irradiation. Seo et al demonstrated an optoelectronic synaptic device with h-BN/WSe 2 van der Waals heterostructure that had both synaptic and optical sensing functions [44]. The synaptic device exhibited good weight update linearity, energy efficiency, and stability of variation. Specifically, it emulated the color pattern recognition capability of human visual systems through an optic-neural network (ONN) formed by optoelectronic synaptic devices. There was over a 90% recognition rate of ONN after 50 training epochs ( figure 3(a)). As the training epoch increased, the synaptic weight was further optimized (figures 3(b) and (c)). ONN achieved recognition of mixed color digit patterns with the highest activation score for complex color mixing numbers.
Simulation results demonstrated the potential of optoelectronic synapses for color recognition. The device-level experiment can further promote the use of neuromorphic vision sensors. Recently, Islam et al presented a multiwavelength in-sensor optoelectronic synapse device, in which monolayer MoS 2 and PtTe 2 /Si played as a conduction layer and the gate electrode for the artificial visual system by performing color recognition, respectively [40]. With infrared light illumination, the electron-hole pairs generated in PtTe 2 /Si modulated the conductivity of the channel through the photogating effect. From UV to visible light irradiation, distinct conductance tuning curves were obtained due to the differences in photoresponsivity ranges of multiwavelength light.
In addition, with the rapid development of electronic eyes and artificial vision systems, the demand for flexible stretchable devices with non-toxic, low power consumption and high robustness is increasing. Recently, Li et al reported a flexible optoelectronic synaptic transistor based on high-quality leadfree CS 3 Bi 2 I 9 NCs and organic semiconductors [45]. The flexible optoelectronic synaptic device exhibited remarkable flexibility and robustness. The prepared flexible synaptic array successfully recognized images of different colors of light under low light power intensity. When 405 nm, 532 nm, and 635 nm light with the same power intensity was irradiated on the device, different EPSCs were generated (figures 3(d)-(g)).
In addition, the current multiwavelength values were distinguishable within the decay time. Therefore, the device can distinguish colors even after the light is turned off. The above results demonstrate the potential of optoelectronic synaptic devices for color recognition.

Angular recognition
A human visual system has the ability of wide-angle recognition and space-time resolution. The eye captures optical information and encodes it into electrical spikes of appropriate size through optoelectronic conversion. Subsequently, the information is transmitted to the visual cortex. According to several factors, such as intensity, environmental angular position and repetition, the information is stored in the natural biological synaptic network, as shown in figure 4(a). The recognition of spatial coordinates and visual orientation in cortical cells is a well-known spatiotemporal processing paradigm, which has been extensively studied in the visual neural network system. The spatial information processing function has been demonstrated using artificial synaptic device arrays.
Recently, Kumar et al reported a simple and effective method for visual processing at a wide viewing angle using Ag NWs/TiO 2 Schottky optoelectronic-coupled photonic devices [49]. Different incidence angles with various light intensity and irradiation times were used to stimuli a 3 × 3 device array ( figure 4(b)). The input image can be well recognized from the post-synaptic current contrast and reliably memorized for a long time. The comparison diagrams showed that the proposed device can perform parallel processing and distinguish spatiotemporal optical input when the illumination angle changed. Therefore, the forgetting behavior of the device can be adjusted according to the lighting angle or intensity. Xie et al reported a 2D neuromorphic device to simulate advanced visual spatiotemporal processing. The 2D MoS 2 electric double-layer transistors with coplanar multigrid input arrays were fabricated [48]. An artificial 2D visual neural network system for simulating the principle of spatiotemporal coordinates and direction recognition was proved through experiments ( figure 4(c)). When the spikes voltage with the same amplitude was applied to different coplanar grid coordinates, different EPSC amplitudes can be obtained due to different gate-channel coupling paths in the MoS 2 transistor. Figure 4(d) summarizes the EPSC triggered by two synchronous spikes of presynaptic terminals. The results showed that the maximum EPSC amplitude corresponds to 0 • visual orientation, while the minimum EPSC amplitude existed at 225 • visual orientation. At the same time, the measured sum and expected arithmetic sum were defined for EPSC amplitude, they can be plotted as a function of presynaptic orientation (figure 4(e)). Similar to the visual system, the coplanar multi-grid input arrays in the proposed MoS 2 transistor were regarded as the receptive field of the visual cortex cells, and the EPSC amplitude was measured as the activity of the spatial cortex cells. The results provide a new way to introduce intelligent visual recognition functions in emerging neuromorphic electronics.

Contrast enhancement
Contrast enhancement is a crucial feature of neuromorphic vision sensors that enables the pre-processing of sensory data, thereby enhancing the quality of sensory information and improving the efficiency and accuracy of subsequent tasks, such as image recognition and classification. Recently, Zhou et al proposed an optoelectronic resistive random access memory (ORRAM) synaptic device with two-terminal structure of Pd/MoO x /ITO, exhibiting non-volatile and volatile  resistance switching upon ultraviolet light sensing and optical triggering, as well as the optically tunable synaptic behavior [52]. The ORRAM arrays demonstrated the real-time image contrast enhancement function. As shown in figure 5(a), the image memorized the letters F and L under different illumination intensity, respectively. Due to the stronger light stimulation, the letter F exhibited a stronger memory effect compared to the letter L. The peak current and retention time of ORRAM devices increased faster with higher light intensity. Accordingly, pixels with higher brightness exhibited a more enhanced cumulative effect. After repeated training, the gray level differences of different pixels were further amplified, resulting in an output image with enhanced contrast. Further, the authors used 3 × 5 ORRAM array to demonstrate the contrast enhancement function of images with four grayscales ( figure 5(b)). Further simulation of the neuromorphic vision system demonstrated that image preprocessing at the front end effectively improved image quality and subsequently improved the efficiency and accuracy of subsequent image recognition. In recent years, optoelectronic devices have been developed to achieve neuromorphic vision sensors by combining optical and electrical signals. However, this approach faced some challenges, including trade-offs between bandwidth, connection density, hardware redundancy, high power consumption and computing delay. Researchers have developed alloptical tunable neuromorphic visual devices to solve these problems. Recently, Ahmed et al reported a neuromorphic visual device based on BP, which had a simple reconfigurable phototransistor structure and inherent all-optical modulation storage capability [159]. Optical tunability was achieved by using oxidation-induced defects in the vertical stack of BP. These defects produced positive and negative photoconductivity under different ultraviolet light irradiation. Based on the bipolar change, the function of image contrast enhancement was realized. A 2 × 2 pixel array was irradiated by the ultraviolet light at both high and low frequencies (figure 5(c)). Compared with the low frequency irradiation, the high frequency irradiation achieved longer conductance memory. The increase in conductivity contrast at higher irradiation frequencies highlighted the device's ability to perform intra-pixel image enhancement. Single-device optoelectronic synapses and finite pixel arrays have shown potential in image sensing, memory, and processing. Wafer-scale vision chips can further promote the development of artificial vision systems. Ma et al designed a wafer-scale 2D MoS 2 vision chip, which integrated 619 pixels with 8582 transistors in an area of 10 square millimeters to achieve the function of contrast enhancement [54]. These results demonstrated the potential for the proposed vision chip to be used in future applications of artificial vision systems.

Motion detection
In the era of the internet, the detection and recognition of mobile objects has become increasingly important. Conventional technology for motion detection and recognition, based on complementary CMOS image sensor platforms, included redundant sensing, transmission conversion, processing, and storage modules. Recently, Zhang et al reported a retinomorphic device based on BP/Al 2 O 3 /WSe 2 /h-BN heterostructure structures, which showed integrated perception, memory and computing capabilities for motion detection and recognition [59]. The non-volatile positive and negative photocurrents can be realized by modulating the channel carrier state through applying the gate voltage, which can simulate the reversible regulation and storage functions of bipolar cells in the retina. The dynamic motion process of an object was considered as a series of pictures at different times (figures 6(a)-(c)). Image brightness distribution at one time (t 1 ) was multiplied by the positive conductivity matrix, and the image brightness distribution at the next time (t 1 + ∆t) was multiplied by the negative conductivity matrix. The stored results at the two times were summed to obtain the processed pixels, and the recognition of moving objects was achieved according to the summed brightness. Compared with conventional motion detection technology, this method greatly reduced the impact of the surrounding environment and minimized the redundant data transmission.
Additionally, Lao et al reported a self-powered optoelectronic synaptic device based on Au/P(VDF-TrFE)/Cs 2 AgBiBr 6 /ITO heterostructure for in-sensor reservoir computing, which realized ultra-low power machine vision applications [57]. Motion detection was demonstrated through photocurrent-encoded in-sensor reservoir computing ( figure 6(d)). The vehicle optical pixel information from the dynamic vision sensor was input into a 5 × 5 photonic synapse array. When five photonic synapses were activated and the EPSC was released, the vehicle direction was recognized. The researchers further designed a 25 × 4 memristor network for the vehicle flow recognition tasks. After 24 training periods, the recognition rate was 100% (figures 6(e) and (f)).

Visual adaptation
Self-adaptive artificial visual system can adapt to environmental changes (such as light and dark changes) to realize object recognition. Recently, Kwon et al reported an array of light-tunable optoelectronic neuromorphic biological devices to simulate the biological visual function, demonstrating the environmental adaptive artificial visual perception system [62]. Figure 7(a) illustrates the concepts of photopic and scotopic adaptation. In human visual adaptation physiology, rod cells are mainly responsible for dark vision, while cone cells are responsible for bright vision. Their combination allows to adapt and perceive a wide range of light intensity. A 3 × 3 pixel array was illuminated with different red light intensity to evaluate the adaptive visual perception ( figure 7(b)). The output current was modulated by the load gate voltage (V L ). Under the high light intensity condition, encoded patterns showed an enhanced contrast at higher V L , similar to the process of photopic adaptation. However, under the same adaptation condition, the pattern disappeared within 1 s when sensing weak light. Therefore, to achieve long-term image retention and weak light discrimination, the V L needed to be reduced to increase the PSC. Through the construction of an optoelectronic neuromorphic device array, the environment-adaptive artificial vision system with light adaptation behaviorwas realized, which greatly improved the efficiency and accuracy of image recognition of the machine vision.
To achieve light-adaptive retinal imitation, it is necessary to construct a flexible artificial vision system. Recently, Meng et al reported a flexible artificial retina perception device based on 2D Janus MoSSe, which can be modulated by the electricity/ion and light [63]. As shown in figure 7(c), the PSC exceeded the threshold value when the device was exposed to the strong light power. The applied negative gate voltage pulses can reduce the PSC to achieve photopic adaptation. Although flexible optoelectronic synaptic devices exhibited excellent bending stability and light adaptation, the previous studies focused only on photopic adaptation instead of time-dependent and light-intensity-dependent dynamic adaptive image recognition. Therefore, Liao et al reported the bio-inspired vision sensor based on the MoS 2 back-gate phototransistor array, which possessed the characteristics of timedependence and light-intensity-dependence, thereby realizing sensing and adaptive functions similar to human retina [65]. The channel was electrostatically doped through the gate voltage to achieve scotopic adaptation and photopic adaptation, respectively (figures 7(d)-(f)). To quantitatively evaluate the potential of visual adaptation to enhance the image recognition, the authors constructed a vision system comprising of the adaptive MoS 2 phototransistor array and a three-layer ANN. By using the modified National Institute of Standards and Technology dataset as the training set to evaluate the recognition accuracy of images in dark and bright environments, the vision system achieved recognition rates of 96.9% and 96.1% in scotopic and photopic adaptation, respectively. These results suggest bioinspired in-sensor potential for wide-range sensing and image recognition.

Conclusion
In this paper, we summarized the research progress of lowdimensional optoelectronic synapses in neuromorphic vision sensors. Firstly, the structure and function of the human retina and optoelectronic synaptic plasticity were introduced, demonstrating how to simulate the visual system at the device level. Secondly, a brief introduction of the typical working mechanisms of optoelectronic synapses based on different low-dimensional materials was reviewed in detail. Finally, we provided a comprehensive overview of the advanced applications of optoelectronic synapses in neuromorphic vision sensors.

Future perspectives
Neuromorphic vision sensors can efficiently collect, memorize, and process external optical information by simulating biological visual systems which have the advantages of high time-domain resolution, low data redundancy, low power consumption, and showing important prospective applications for the development of the internet. The basic units of neuromorphic vision sensors are optoelectronic synaptic devices that achieve dual functions of photosensitivity and synapses mainly through light excitation. Most of the existing neuromorphic vision sensors are based on mature silicon-based CMOS technology, such as the silicon retina. However, conventional neuromorphic vision sensors are mainly built on rigid substrates and typically only perform demonstration and performance optimization of neuromorphic vision, making it difficult to construct a curved vision system similar to the shape of the human eye. At the same time, the circuit of silicon-based neuromorphic vision devices is complex with large pixel area and low fill factor, leading to low responsiveness and poor uniformity under high-density integration. In addition, the inherent von Neumann architecture of siliconbased neuromorphic visual sensors makes them inefficient to match biological visual systems. Fortunately, the construction of bio-inspired low-dimensional optoelectronic synapses has injected new vitality into the development of neuromorphic vision sensors. Compared with conventional neuromorphic vision sensors, new devices based on low-dimensional materials exhibit various excellent performances, such as a strong light-matter interaction, rich defect states, and flexibility. In addition, heterogeneous integration of multiple materials enables the customized design of band structures, rendering them ideal platforms for neuromorphic vision sensors. In recent years, advanced applications of neuromorphic vision sensors based on low-dimensional materials have developed rapidly. In terms of visual information perception, it has achieved spatial angle perception, dynamic motion detection, and self-adaptive functions. In terms of visual information processing, image contrast enhancement and color recognition have been achieved. Even though the neuromorphic vision sensors based on low-dimensional optoelectronic synapses faced many challenges in achieving or surpassing the performance of the biological vision, for example, the compatibility between the weight enhancement and suppression under different optical stimuli, the symmetric weight update, and the development of hemispherical visual systems. In addition, in terms of devices, most current reports focus on simulating the synaptic function which can only perform the static simulations. However, it is particularly important to achieve synaptic plasticity under dynamic regulation for optoelectronic synaptic devices. In addition, establishing standards to evaluate device performance can help explore superior materials or structures for neuromorphic vision sensors. Besides, most current studies of optoelectronic synapses based on lowdimensional materials focus on the single device or small-scale arrays restricted by the material ′ s growth. Therefore, largescale low-dimensional material ′ s synthesis with high uniform quality and low cost can provide an ideal platform for constructing neuromorphic vision sensors with high integration and consistency. In terms of algorithms, it is complicated to process visual information similar to the human vision system and achieve low-power and efficient parallel computing. Therefore, developing algorithms based on new materials and devices is also an urgent task. In the future, how to fully utilize low dimensional optoelectronic devices to achieve high integration, wide field of view, low power consumption, and low latency artificial vision systems will be an important research direction.