The sustainable development of conventional computers with von Neumann architecture and silicon complementary metal–oxide semiconductor (CMOS)-based hardware has been hampered as follows: (i) devices restricted by scaling theory near the limits of physics; (ii) von Neumann bottleneck due to physically separated storage and data processing units [1–4]. As the highest structure of the nervous system, the human brain is responsible for processing information received by various senses, an activity that depends on the existence of 100 billion neurons and 1000 trillion synapses interconnected in the cerebral cortex at an ultra-high density [5–7]. In contrast, the human brain can execute complex tasks such as parallel computation and cognitive learning with the advantages of being highly fault-tolerant and event-driven, thus generating interest in brain-like computers [8–10]. Over the past few years, significant progress has been achieved regarding brain-like chips, among which the best-known one is the TrueNorth chip introduced by IBM in 2014 [11–13]. However, since such chips for high-speed computation still suffer from low integration density and high energy consumption compared with the human brain due to the limitations of the traditional CMOS structure, the emergence of synaptic electronics will lead to the future of artificial intelligence (AI).

Inspired by biological synapses, several emerging devices with the advantages of simple device structure and high-density integration, such as phase-change memories, resistive switching memories, and field effect transistors, were proposed to mimic the synaptic plasticity [14–19]. To date, our comprehension concerning the complex characteristics of synaptic devices remains at a preliminary level, with a growing appreciation of the essential role played by both material selection and signal modulation. Despite enormous efforts that have been devoted to investigating electrical synapses, the all-electronic design raises issues of high energy consumption, limiting further applications in artificial neural networks (ANNs) [20]. Here, taking into account the important role played by artificial vision systems (AVSs) in neuromorphic engineering, optoelectronic synapses exhibiting excellent characteristics in terms of large bandwidth and low power consumption have been developed as the foundation for the next generation of neuromorphic systems [21,22]. So far, metal oxide films, organic semiconductors, and other materials have been successively employed in synaptic devices modulated by light signals or photoelectric synergistic properties [23–28].

Metal halide perovskites (MHPs) have emerged as revolutionary photosensitive materials with the generic formula ABX3, in which A represents the cation, B denotes the divalent metal ion, and X refers to the halide [29–31]. In the past few years, the family of MHP materials has drawn considerable attention for being employed in a wide range of electronic devices like solar cells, light-emitting diodes, photodetectors, phototransistors, memory devices, flexible devices, and lasers, attributed to their superior combination of optoelectronic properties, including large light absorption coefficients, long carrier lifetimes/diffusion lengths, decent charge carrier mobility, and low exciton binding energies [32–45]. Interestingly, compared to conventional organic and inorganic materials, electronic devices employing MHPs tend to exhibit ubiquitous hysteresis effects, a phenomenon that motivates MHPs materials to serve as the most desirable contenders for simulating biological synapses, typically attributed to the inherent ion migration properties or charge carrier traps of MHPs [46–50]. To date, multiple types of optoelectronic synaptic devices based on MHPs have been developed, leveraging the dependence in synaptic plasticity upon optical pulses, and they hold the promise of constructing multifunctional artificial neuromorphic systems that perceive external environmental changes as well as the processing of performance information, in addition to accomplishing brain-like computational behaviors for higher-order learning and handwritten digit recognition [51–53].

In this manuscript, the current status and applications of MHP-based optoelectronic synapses during recent years are comprehensively summarized in Fig. 1 [23,54–60]. A brief description of the synaptic functions of the optoelectronic synaptic devices is presented in Section 2, as well as the calculation regarding the energy consumption involved in the synaptic events. Focusing on device architecture, Section 3 discusses in detail MHP-based optoelectronic synapses from the perspectives of both structural design and potential physical mechanisms, providing innovative approaches for the exploitation of the novel artificial synapses. Subsequently, the innovative functional neuromorphic applications utilizing MHP-based synaptic devices for associative learning, emotional learning, logical functions, arithmetic operations, and neuromorphic visual systems (NVSs) are further reviewed in Section 4. In conclusion, the obstacles and challenges surrounding the optoelectronic synapse are outlined in Section 5, along with reasonable predictions related to their applications.

Synapses, as an essential component of information transmission throughout the nervous system, provide a suitable imitation object for researchers to build ANNs. In recent years, more and more MHP-based optoelectronic synaptic devices have been proposed to mimic the basic functions of biological synapses, such as synaptic plasticity [51,57,61,62]. In Section 2, the connection between neurons and synapses is described, followed by a highlight of the typical synaptic behavior of MHP-based optoelectronic synaptic devices and their performance metrics. How to achieve these synaptic behaviors with as little energy consumption as possible remains a great challenge for the construction of the devices [63–66].

Neurons, also called nerve cells, are the most basic structural and functional units of the biological nervous system. Synapses are located between two neurons that are in contact with each other and assume the function of transmitting information in a unidirectional manner within the nervous system. There are two types of biological synapses: chemical synapses and electrical synapses, which use chemical and electrical signals to transmit information, respectively. In this work, we focus on the working mechanism of chemical synapses. As shown in Fig. 2(a), a synapse consists of a combination of a presynaptic membrane, a postsynaptic membrane, and a narrow space between them (the synaptic cleft).

When a nerve impulse is transmitted to the presynaptic membrane, voltage-gated calcium ion channels in the membrane open, and extracellular Ca2+ enters the presynaptic element, allowing neurotransmitters to be released into the synaptic cleft [70]. Some neurotransmitters bind to the corresponding receptors on the postsynaptic membrane, resulting in the opening of chemically gated channels through which the corresponding ions enter the postsynaptic element [71]. If excitatory neurotransmitters are released from the presynaptic membrane, which increases the permeability of the postsynaptic membrane to Na+ and K+ increases, resulting in a localized depolarization potential—that is, an excitatory postsynaptic potential (EPSP)—then the synapse is referred to as an excitatory synapse. Conversely, if the presynaptic membrane releases inhibitory neurotransmitters that increase the permeability of the postsynaptic membrane to Cl−, leading to a local hyperpolarization potential, or inhibitory postsynaptic potential (IPSP), then the synapse is called an inhibitory synapse. It is worth noting that there are many synapses in a neuron. When the sum of the excitatory synaptic activity in the neuron exceeds the sum of the inhibitory synaptic activity and action potentials in the axons of the neuron are triggered, leading to the occurrence of nerve impulses, the neuron is presented as excitatory and, conversely, as inhibitory.

During this dynamic process of neurons transmitting information, the strength of the connection between them is determined by the synaptic weight [68,72]. For biological nervous systems, variations in synaptic weights are defined as synaptic plasticity, which causes postsynaptic currents (PSCs) to fluctuate in response to the activity of presynaptic neurons and is regarded as fundamental to the human brain’s ability to recognize, encode, store memories, and discriminate information [73,74].

For synaptic devices, external stimuli such as optical/electrical pulses applied to the electrodes correspond to the release of nerve impulses in biological synapses. The channel conductance is equivalent to the synaptic weight, and the resulting channel currents are referred to as excitatory postsynaptic current (EPSC) or inhibitory postsynaptic current (IPSC), corresponding to EPSP and IPSP of biological synapses, respectively. In general, the value of the PSC is determined by the synaptic weights. As the most fundamental synaptic behavior that can be observed in optoelectronic synaptic devices, PSC is regarded as an essential component of ANNs. As shown in Figs. 2(b) and 2(c), PSC occurs when the perovskite-gated synaptic device is stimulated by a single light pulse. Here, EPSC or IPSC is controlled by the gate voltage of the synaptic device [67]. In addition to the modulation of the gate voltage, the wavelength of the optical stimulus is also significant in affecting the PSC [60,75]. Hou et al. reported an ambipolar optoelectronic synaptic device based on a Pyr-GDY/graphene/PbS quantum dot (Pyr-GDY/Gr/PbS-QD) heterojunction. As shown in Fig. 2(d), the triggered IPSC is converted to EPSC when the wavelength of the incident optical pulse is switched from 450 to 980 nm [75]. Furthermore, for synaptic devices, the magnitude of the PSC is determined by the strength and duration of the applied light/electrical pulse, as shown in Figs. 2(e) and 2(f) [23]. The decay of PSC with time after the pulse applied to the device has disappeared can be perfectly fitted by the Kohlrausch stretched exponential function (SEF) [76–80], I=(I0−I∞)exp[−(t−t0τ)β]+I∞,(1)where I0 and t0 stand for the PSC and time when the pulse stimulation ceased, respectively; I∞ denotes the value at the final stabilization of the current; and τ and β refer to the retention time and the stretch index in the range 0–1, respectively. Figure 2(g) shows a typical decay curve of EPSC with time [69], from which it can be seen that the value of EPSC decreased first and then tended to be stable, closely matching the curve fitted according to the SEF.

A pair of pulse stimuli is applied to the synapse, as shown in Fig. 3(a), which modifies the concentration of Na+/Mg+ and other ions that contribute to the change in the postsynaptic membrane potential, which in turn affects the magnitude of the PSC after the two stimulations [54], causing a variation in the degree of connectivity between the two adjacent neurons [72]. Similarly, if two consecutive pulses of stimulation are applied to the synaptic device, the spikes of PSCs obtained from the first and second stimuli are denoted as A1 and A2, respectively. If the value of A2 is larger than A1, which implies an enhanced postsynaptic response, then this synaptic behavior is described as paired-pulse facilitation (PPF); conversely, the phenomenon indicates a depressed postsynaptic response, as well defined as paired-pulse depression (PPD) [83]. The dependence of PPF/PPD behavior on the pulse interval time makes it assume an important role in decoding the temporal data in the synaptic signal [84,85]. Wang et al. proposed a synaptic device based on MoS2/PTCDA heterostructure and obtained typical PPF/PPD behavior by applying a pair of gate pulse stimuli to it [Figs. 3(b) and 3(c)] [81]. The PPF/PPD index can be calculated based on the classical equation (A2/A1)×100% or [(A2−A1)/A1]×100% to calculate [54,55]. Interestingly, the facilitation/depression percentage gradually decays/strengthens to 100% with the increase of pulse interval time, and the attenuation/enhancement curve can be well fitted by the following double-exponential function [Figs. 3(d) and 3(e)] [21,54,68,74,75,84]: PPFindex=1+C1exp(−Δt/τ1)+C2exp(−Δt/τ2),(2)where C1 and C2 are defined as the initial facilitation constants of the biological synapse; τ1 and τ2 are considered as fast and slow characteristic relaxation times, respectively, which can be extracted from the fitted curves; and Δt is the interval time between two consecutive pulse stimuli. Notably, in biological synapses, the values of τ1 and τ2 both range from milliseconds to seconds, where the value of τ1 tends to be an order of magnitude smaller than the value of τ2 [68,72,81]. Furthermore, it has been reported in the literature that PPF behavior is also influenced by numerous external factors, for instance, the wavelength of the optical stimulus [Figs. 3(f) and 3(g)] [1,55] or the spikes of the electrical stimulus [Fig. 3(h)] [82]. Similarly, the mechanisms of PPF behavior in the neurosynaptic devices are caused by many factors, such as the growing number of conductive ions accumulated at the interface when τ is greater than Δt, allowing the channel conductance to be enhanced with the increase of the pulse number [86].

Depending on the duration of changes in synaptic weight, synaptic plasticity is divided into short-term plasticity (STP), which gradually recovers to its initial state within a short period after being triggered, and long-term plasticity (LTP), which maintains variations in synaptic weight for minutes, weeks, or even permanently [68,72]. The above-mentioned EPSC, IPSC, PPF, and PPD are all typical manifestations of STP, which is stored in the hippocampus of the brain according to the model of learning and memory proposed by Atkinson and Shiffrin [Fig. 4(a)] [87,89]. After a constant process of consolidation and rehearsal, the synaptic weights are strengthened, and STP transforms into LTP for storage in the cerebral cortex [68,89,90]. In the human brain, the STP assumes the responsibility of receiving and transmitting information, while LTP, as a typical representative of Hebb’s plasticity, is regarded as the basis for achieving learning and memory functions [91–93]. Referring to Hebb’s postulate (this synaptic efficacy improves if the postsynaptic neuron is stimulated successively and repeatedly by the presynaptic neuron) [94,95], the synaptic weight can be adapted by varying the pulse wavelength, power density, pulse width, and number of incident light pulses to implement the transition from STP to LTP further [Figs. 4(b)–4(e)], which is the basis of the learning behavior imitated by synaptic devices [85,87,91,96–98]. Notably, following the enhancement or diminution of synaptic weight, short-term potential (STP) and long-term potential (LTP) are classified as short-term depression (STD) and long-term depression (LTD), respectively [99].

Spike-dependent plasticity (SDP) is regarded as the dependence of synaptic plasticity on spiking stimuli, with specific manifestations of various types such as synaptic duration-dependent plasticity (SDDP), spike-rate-dependent plasticity (SRDP), synaptic voltage-dependent plasticity (SVDP), and spike-number-dependent plasticity (SNDP), which reflect the implications of the width, frequency, amplitude, and the number of spike pulses transmitted from presynaptic neurons on synaptic weights, respectively [51,90,100–105]. In terms of SRDP, two features of LTP, known as LTP and LTD, are triggered by pulse spikes at high frequency from 20 to 100 Hz and low frequency from 1 to 5 Hz, respectively [93,106].

Similar to SRDP, spike-timing-dependent plasticity (STDP) is one of the representative forms of LTP, which was first initially proposed as the computer learning algorithm widely applied in machine intelligence and now reflects the dependence of synaptic weights (ΔW) on the sequence and time interval (Δt, Δt=tpost−tpre) of pre- and postsynaptic neural activities in biological synapses [107–110]. More specifically, when the presynaptic spike arrives ahead of the postsynaptic spike (Δt>0, prepost-pairing), the pulse signal induces the generation of LTP, with an increase in synaptic weight; on the contrary, the degree of synaptic connection diminishes at Δt<0 (postpre-pairing), and LTD behavior is triggered [92,95,111,112]. Various spiking patterns and relative spike durations influence the direction and amplitude of synapses, inducing the generation of the typical four forms of STDP, namely the antisymmetric Hebbian learning rule, antisymmetric anti-Hebbian learning rule, symmetric Hebbian learning rule, and symmetric anti-Hebbian learning rule, respectively [56,88,94,113,114]. Li et al. reported a GST-based memristor synaptic device relying on the capturing and releasing of charge traps caused by defects within the material itself, and four different forms of STDP can be well exhibited by the device, as shown in Figs. 4(f)–4(i) [88]. In computational neuroscience, asymmetric STDP learning rules can be modeled by exponential functions, while symmetric STDP learning rules are fitted by Gaussian functions, which can be summarized as [64,88,93,115] ΔW={Ae(−Δtτ)+W0,asymmetricAe(−Δt2τ2)+W0,symmetric,(3)where A and W0 are defined as the scaling factor of the function and the constant indicating the non-associative component for synaptic weight alteration, respectively, and τ is considered as the time constant of the fitted function.

The energy consumption generated when any synaptic behavior has been triggered is considered one of the essential indicators to evaluate the performance of synaptic devices, and the calculation was proposed by Kuzum et al. in 2013 [116]. Specifically, the energy consumption E=V×I×t, where V is defined as the electrical pulse amplitude, t is defined as the width of the electrical pulse, and I is regarded as the triggered PSC [117]. In contrast to electrical synapses, a new calculation method of energy consumption was introduced by Tan et al. for optical synaptic devices: E=P×S×t, where P is determined as the power density of the incident light pulse, t is defined as the optical pulse width, and S is defined as the effective illumination area of the synaptic device [118]. Remarkably, one of the two energy consumption calculation methods above applies to synaptic events triggered by electrical pulses, while the other focuses on the energy consumption induced by optical pulses. When the synaptic behavior is stimulated by optical and electrical signals in parallel, then the energy consumption is also obtained by superimposing the two calculation methods above [63].

The human brain consumes a fixed amount of energy to carry out the normal physiological activity, and the cost of neuronal release spikes is so prohibitive that only no more than 1% of neurons are active simultaneously [119]. Furthermore, in biological nervous systems, 1–10 fJ energy is consumed per synaptic spike, a metric that can be accomplished by the few synaptic devices proposed in recent years [116,120]. However, the energy consumption of most synaptic devices is quantified in pJ or even nJ, which hinders the development of neuromorphic computing [63,64,121–124]. To achieve large-scale integration of synaptic devices, the limitation of high energy consumption must be overcome, which can be optimized by weakening the programming pulse amplitude, shortening the programming pulse width, decreasing the effective area of the synaptic device, and reducing the programming current, respectively [65,66,125].

To date, significant efforts have been devoted to designing and fabricating MHP-based artificial synaptic devices for achieving synaptic plasticity. According to a different stimulation source, MHP-based optoelectronic synapses can be divided into all-optical stimulated synapses and optically–electrically synergistic stimulated synapses. In this section, we focus on two-terminal memristors and three-terminal transistors employed for artificial optoelectronic synapses, as well as analyzing their device architectures and operational mechanisms. More specifically, the fundamental synaptic characteristics (terminal number, synaptic functions, energy consumption, etc.) are summarized in Table 1.Table 1.

Summary of MHP-Based Optoelectronic Synapses

Two-terminal memristors, first proposed by Prof. Chua in 1971 [140], are widely employed in hardware implementation for neural networks due to advantages such as a simple fabrication process, low energy consumption, and large-scale integration [141,142]. Interestingly, for such memristors with inherent switching effects, the transition between multiple resistive states relies on the electrical-bias-history-dependent resistance, a characteristic applicable to simulating synaptic behaviors [143–145]. A typical MHP-based optoelectronic synaptic memristor with an Ag/MAPbI3/Ag horizontal structure was presented by Zhu and Lu [Fig. 5(a)] [76]. As shown in Fig. 5(b), resistance switching (RS) dynamics can be tuned by optical stimulation since illumination inhibits the formation of iodine vacancy (VI˙/VI*) in MAPbI3 while simultaneously facilitating its spontaneous annihilation process. In addition, the accumulation and spontaneous decay of the concentration in VI˙/VI* are intimately related to the stimulus of the input pulses, and the MAPbI3-based memristor can mimic synaptic behavior, such as STP and LTP benefiting from the similarity of this dynamics of VI˙/VI* to Ca2+ of biological synapses. Further, the dependence of the spontaneous attenuation rate of the conductance upon illumination and the LTP/LTD behavior of the device is illustrated in Figs. 5(c) and 5(d), implying that the synaptic behaviors of the MAPbI3-based memristor are synergistically regulated by the optical and electrical stimuli. Recently, the Au/KI-MAPbI3/ITO optoelectronic synaptic memristor with vertical structure was proposed by Lao et al. utilizing the excellent optical properties of MAPbI3 [130]. The device architecture is shown in Fig. 5(e), where the additive KI was introduced to enhance the crystallinity of MAPbI3 while modifying its surface defects, which facilitated the ion migration of the mixed film, in turn leading to the dynamic range of the memristor conductance being further extended. In this case, the conductance level initially increases sharply after the application of 800 consecutive voltage pulses followed by a steady level to the maximum value, while the conductance relies on the illumination assuming a synergistic stimulus role, meaning that the memory level of the electrical input is enhanced upon illumination [Fig. 5(f)]. Further, the classical LTP and LTD behaviors triggered by the voltage stimuli applied under illumination were observed [Fig. 5(g)].

Compared with conventional electrically driven synaptic devices, light-driven synapses have attracted much attention due to lower interconnection energy consumption, faster signal transmission, and higher bandwidth [29]. Given this, Yang et al. demonstrated an all-optical two-terminal synaptic memristor with an ITO/SnO2/CsPbCl3/TAPC/TAPC:MoO3/MoO3/Ag/MoO3 vertical structure [Fig. 5(h)] [139], where the TAPC layer serves as a hole transporting layer and the CsPbCl3 film layer assumes the responsibility of a UV light absorber. Eventually, based on the working mechanism of UV-induced surface charge trapping/detrapping under the stimulus of optical spikes, this dual-mode tuned all-optical synapse can well accomplish biological synaptic behaviors such as PPF, SNDP, and SFDP [Figs. 5(i)–5(k)].

Unlike two-terminal memristors, three-terminal transistors have multi-gate structures comparable to biological dendrites, facilitating synergistic modulation of input pulses, stability of operation, and the transmission of synaptic signals [139,146–148]. Emerging MHP-based synapse phototransistors (SPTs) have attracted attention due to their outstanding photovoltaic conversion efficiency, further demonstrating their structural stability and operational controllability. For instance, Park et al. have designed an ITZO/BA2PbBr4 SPT for NVSs with the structure shown in Fig. 6(a), which regulates synaptic behaviors by coupling stimulation with optical and electrical spikes [128]. When the light signal was applied to the BA2PbBr4-based SPT, the photogating effect was induced due to the regulation of the energy band at the interface between ITZO and BA2PbBr4 facilitating the separation of photo-excited carriers, thereby increasing the value of PSC by capacitive coupling [Fig. 6(b)]. Additionally, the characteristics of LTP and LTD were observed, which are generated by the photogating effect triggered by 50 successive light pulses (100μW/cm2, 1 s) and the recombination of photo-excited carriers prompted by 50 successive positive gate pulses (20 V, 2 s) [Fig. 6(c)]. Notably, the value of EPSCs in LTP characteristics increased with the decay of λ, following the optical absorption spectra of the BA2PbBr4 film. Not only the gate voltage pulses but also the drain electrical spikes of the SPTs are available for the regulation of the conductance state as demonstrated in a graphene-PQD (G-PQD) superstructure exploited by Pradhan et al. [138]. The current-voltage characteristic of the proposed structure with and without illumination is shown in Fig. 6(d), where quantum dots act as the photoabsorbing material and graphene assumes the responsibility of an effective carrier transport channel. The behaviors of LTP and LTD can be achieved as band-to-band/impurity-to-band transitions in the SPT prompting an effective charge transfer from PQDs to the graphene layer. Moreover, the study briefly analyzed the gate-dependent LTP transient characteristic of the G-PQD SPT stimulated by 20 consecutive light spikes (1.1μW/cm2, 5 s) [Fig. 6(f)].

Huang et al. broadened the absorption spectrum of PEA2SnI4 perovskite by inserting a Y6 organic film, enabling the proposed PEA2SnI4/Y6 ambipolar SPT to actualize the red/green/blue/near-infrared (NIR) wavelength selectivity [60]. With the structure illustrated in Fig. 6(g), this SPT accomplishes synaptic plasticity behaviors in response to stimuli of optical spikes, relying on charge trapping/detrapping processes at the PEA2SnI4/Y6 interface. Compared with the optically–electrically synergistic stimulated devices, as an all-optical type device, the advantage of this structure is that the increase and decrease of conductance can be modulated by the optical signal only, contributing to avoiding the generation of joule heat during the operation. In this work, an IPSC/EPSC was triggered when the SPT was stimulated by a visible/NIR light spike [Fig. 6(h)], with the mechanism of operation shown in Fig. 6(i). Here, electron–hole pairs were generated in both the PEA2SnI4 film and Y6 under the stimulus of visible optical spikes, where photogenerated electrons were trapped by Sn vacancies present in the PEA2SnI4, which resulted in an enhanced photogating effect and evoked more holes to recombine with partial electrons in the channel, leading to a reduction in the electron concentration and PSC values of the SPT eventually. As opposed to this, when an NIR light spike was applied, benefiting from the generation of electron–hole pairs in Y6, the holes outflowing from Y6 were trapped at the PEA2SnI4/Y6 interface, which further contributed to the increase in the values of the PSCs.

Recently, MHP-based optoelectronic synaptic devices have succeeded in mimicking the basic synaptic functions by skillfully incorporating light as the pulse signal, which builds a bridge to construct AI and brain-like computing at the physical level. Further, this section summarizes the novel application perspectives of the devices and classifies them roughly into three key categories: (i) neuromorphic computing, (ii) high-order learning behaviors, and (iii) NVSs.

Due to the increasing demand for storage and computational accuracy, most resistive random-access memories (RRAMs) that specialize in binary storage face the challenge of insufficient reproducibility, attracting attention to innovative computational concepts [2,149–151]. In fact, since the increased/decreased MHP-based optoelectronic device currents in response to optical/electrical impulses strongly resemble the addition/subtraction calculation of the abacus, the proposed MHP-based optoelectronic synaptic devices on this basis can serve to accomplish decimal arithmetic functions, which have implications for the implementation of algorithms other than binary [152–154]. Huang et al. reported the MHP-based optoelectronic memristor with MAPbI3/SiNCs hybrid film fabricated by a two-step method in 2020 [20]. Figure 7(a) illustrates the increase of the initial current (EPSCi) as a function of the stimulus number (i) of the device under optical spikes irritation (with interval time of 0.2 s). In general, a perfect linear correlation has to be exhibited between i and ESPCi to satisfy the arithmetic requirement of decimal counting (here, the linear fitting Pearson correlation coefficient is up to 0.99) [155]. When m+n is calculated, the value of i is derived by matching the exciting current value with the specific EPSCi by adding n consecutive pulse stimuli after m consecutive pulse stimuli. The operation of subtraction works in the same way as addition [151], and the value of m−n can be calculated by computing the amount of optical excitation necessary for the EPSCi change from EPSCn to EPSCm [Fig. 7(b)]. In the case of “9 + 7,” the EPSC excited by 9 plus 7 consecutive optical pulses reaches 162 pA, which equals the value of EPSC16, corresponding to the calculation process of “9 + 7” = 16 [Fig. 7(c)]. Apart from simple addition algorithm, the operation of subtraction, multiplication, and division can also be implemented by these MHP-based synaptic devices, as shown in [Figs. 7(d)–7(f)].

As is well known, not only arithmetic computation but also dynamic logic operations play an important role in reflecting the data processing capability of the neural system [150], providing additional support for achieving multifunctional neuromorphic computing [154,156,157]. Based on multiple-light-stimulated synapse of DPPDTT/CsPbBr3 QDs with the capability of mimicking essential synaptic behavior [Fig. 8(a)], two series of optical signals at 450 and 500 nm were employed by Hao et al. to implement Boolean logic arithmetic including both “AND” and “OR” logic functions [158]. Much more interestingly, at a weak optical illumination of 0.12mW/cm2, the ΔEPSC value increases beyond a threshold current (5 nA) only when two peaks are imposed on the synapse separately, illustrating the “AND” logic function [Fig. 8(b)]. Nevertheless, with an increased irradiation density of 0.3mW/cm2, the ΔEPSC triggered by either stimulus remains high compared to the threshold current, signifying the implementation of the “OR” logic function [Fig. 8(c)].

In addition to simple “AND” and “OR” logic operations, other dynamic logical functions such as the nervous system’s ability to process information can be embodied by MHP-based optoelectronic synaptic devices [53,136]. For example, Zhang et al. exploited optoelectronic transistors to achieve synaptic functions [Fig. 8(d)] [56], in which the CsPbBr3 QDs/DPP-DTT heterostructure facilitates photonic modulation, while the electrical modulation arises from the electric double layer effect in the ionic conductive cellulose nanopaper (ICCN). Here, electrical and optical signals work in tandem to control the switching of logical functions during equipment operation. The output current is above the threshold value (−60pA) only when two electrical spikes (−0.5V) are stimulated simultaneously without light modulation, corresponding to the “AND” logic function [Fig. 8(e)], while the current triggered independently by any electrical spike exceeds the threshold current when modulated by weak light and is considered as the “OR” logic operation [Fig. 8(f)]. Further, the “NAND” logic operation modulated by medium light is proposed when both electrical spikes added to the gate electrode are 0.3 V [Fig. 8(g)], and the “NOR” logic operation is induced by raising the light intensity to strong light [Fig. 8(h)].

It is well known that brain-inspired neuromorphic computing breaks the limitations of computing speed, while the emergence of novel synapse-based ANNs provides the potential to render their terminal form [56,159,160]. Han et al. designed a light-stimulated synaptic transistor device based on a graphene/hexagonal boron nitride (h-BN)/perovskite QD structure and then investigated the learning capability of the device through the Modified National Institute of Standards and Technology (MNIST) database [57]. As shown in Figs. 9(a) and 9(b), the device was stimulated with 100 continuous light pulses and 100 continuous electrical pulses to obtain 200 conductance values, which were used as synaptic weights by the researchers to construct a supervised learning framework for ANN; 784 input neurons, 300 hidden neurons, and 10 output neurons were linked by synaptic weights to form the ANN [Fig. 9(d)]. The recognition accuracy of the device-based ANN gradually saturates with the amount of training, and when tested with handwritten digit images from the MINIST dataset, the overall recognition accuracy of up to 91.5% can be seen at around 40 training epochs [Fig. 9(c)]. The authors then directly demonstrate the superior recognition ability of the ANN by showing the weight mapping matrix of the output images. As shown in Figs. 9(e) and 9(f), the originally ambiguous outlines eventually present the numbers clearly after continuous training, which shows that the ANN has excellent pattern recognition ability and provides another possibility for the development of neuromorphic computing.

As a fundamental characteristic of event-driven learning in neural network systems [161], STDP learning rules are increasingly proposed as a basis for neuromorphic modeling built on CMOS circuits or memristor synapses [56,162–169]. Yang et al. proposed a synaptic memory with an ITO/CsPbBr2I/P3HT/Ag structure, which benefited from the migration properties of halogen ions and succeeded in imitating the STDP behavior of synapses [133]. As shown in Fig. 10(a), a series of presynaptic pulses (Vpre, amplitude of 0.7 V) and postsynaptic pulses (Vpost, amplitude of −0.3V) with a pulse width of 1 s were applied to the Ag electrode and ITO electrode at both terminals of the memristor, respectively. Here, synaptic weights (ΔW) reach a maximum when both Vpre and Vpost arrive at the same time, followed by a decline with increased time interval (Δt), denoted as the symmetric Hebbian learning rule [Fig. 10(b)]. When reversed spike Vpre (amplitude of −0.7V) and Vpost (amplitude of 0.3 V) stimuli are applied to the memristor, the synaptic weights change from excitatory to inhibitory previously, also diminishing reliance on the increasing Δt, and in this way, the symmetric anti-Hebbian learning rule is observed [Fig. 10(c)]. In biological systems, the majority of the information is available through vision. As well for synaptic devices, the introduction of optical pulses as stimulus signals to mimic STDP holds enormous promise apart from electrical pulse activation. In Fig. 10(d), silicon solar cells are connected in series with the ITO electrode and Ag electrode of the memristor for constituting the artificial retina system. During the operation of the device, the photovoltaic cell generated electrical spikes upon two sets of optical pulses, which corresponded to the Vpre (amplitude of 50mA/cm2, width of 30 s) and Vpost applied to the ITO and Ag electrodes, respectively. For the solar cell, if the anode and cathode are connected to the ITO electrode and the Ag electrode, ΔW will be decreased with the increased Δt between the two sets of light optical pulses, and the symmetric Hebbian learning rule will be demonstrated [Fig. 10(e)]. In turn, when the cathode and anode of the solar cell were connected to Ag and ITO electrodes, respectively, the inhibitory ΔW attenuated depended on the growing Δt as well, implementing the symmetric anti-Hebbian learning rule [Fig. 10(f)].

Associative learning refers to the influence of the relationships between different events on the brain’s learning, inspired by the brain’s associations [170,171]. Pavlov’s dog experiment, also known as the Pavlovian conditioned reflex, first proposed in 1927, has been simulated by many synaptic devices as a classic case of associative learning [53,130,136,172–175]. Cheng et al. reported a phototransistor based on the CsPbBr3−QDs/MoS2 mixed-dimensional vertical van der Waals heterojunction (MVVH), which can mimic classical conditioning utilizing photoelectronic hybrid inputs [58]. Herein, food and bell ringing are identified as an unconditioned stimulus (US) and neutral stimulus (NS), respectively, while salivation is defined as an unconditioned response (UR). At first, the UR could be activated by the US alone, but after repeated stimulation by forming an association between US and NS, a separate NS can activate UR, which means that the NS is transformed into a conditioned stimulus (CS) [Fig. 10(g)] [58]. The specific experimental procedure is shown in Fig. 10(h), where 10 consecutive electrical spikes (amplitude of 0.13 V, width of 10 ms) and optical spikes (amplitude of 225mW/cm2, width of 10 ms) are employed to simulate NS and US, respectively. Herein, the EPSCs triggered by the electrical pulse stimulus individually failed to reach the threshold current (ITH), implicating that the dog was unable to generate salivation. After 10 optical and electrical pulse stimuli were applied in parallel, the EPSCs were found to exceed ITH when stimuli with electrical pulses were imposed alone, meaning that the synergetic training conferred the ability to induce conditioned response (CR) in NS. Finally, the EPSC gradually decayed to the original state after a period of US, which indicates the disappearance of the association between US and NS, corresponding to the elimination of redundant information in the human brain [176,177]. Moreover, the triggered EPSCs were significantly increased by increasing the photoelectric synergy training time, as shown in Fig. 10(i), reflecting the fact that, in the case of associative learning, the degree of learning and memory is influenced by the intensity of training. As a further manifestation of associative learning, the reward and punishment mechanism refers to the fact that creatures can be trained by reward and punishment to seek benefits and avoid harm when making decisions. Duan et al. introduced IGZO/PVK NPs/IGZO heterostructure-based synaptic devices to mimic the photoelectric-synergistically typical reward and punishment mechanism [Fig. 10(j)] [131]. In this case, opening windows (NS), feeding (reward), and punishment are simulated by optical pulses, negative voltage pulses, and positive voltage pulses, respectively. As a result, after synergistic training with NS and punishment/reward, the triggered EPSCs (biological responses) remained below/above the threshold value even when the punishment/reward behavior was applied individually [Figs. 10(k) and 10(l)].

It is well known that emotion, as an emotional experience or response, plays a critical role in learning effectiveness; however, the complex feedback mechanism linking learning and emotion has not been completely elucidated by any research [178–180]. In recent years, emulating this biological behavior with synaptic devices has attracted the interest of researchers [181–183]. Yin et al. reported a SiNM/MAPbI3 heterostructure-based optoelectronic synapse that simulates visual learning and memory behaviors, depending on various emotional states [23]. As shown in Fig. 11(a), 30 consecutive optical pulses (wavelength of 532 nm, amplitude of 1μW/cm2, width of 200 ms) are applied to the synaptic device acting as the learning process, while a maximum value of the triggered EPSC is defined as the learning result, and the decay of EPSC over time is considered as the forgetting process [Fig. 11(b)]. Further, the positive, neutral, and negative moods correspond to the situations when the gate voltage is equal to −4V, −3V, and 1 V, respectively. As a result, the process of learning training for the letter “H” in three different mood states is shown in Fig. 11(c). When people learn with positive moods, the impression of the letter “H” is clearest as well as the most effective; as the mood state turns from neutral to negative, the impression of the letter “H” also changes from general to ambiguous, with weaker learning effects. Notably, the retention of memory after tens of seconds under different moods is similar to the above. In addition to electrical signals, optical signals can also be employed to mimic mood states. For example, Lao et al. demonstrated a device based on the Au/KI−MAPbI3/ITO vertical structure by inserting potassium iodide (KI) into MAPbI3 film, which can mimic the learning and memory behaviors that occur when the brain is exposed to positive/negative mood states [130]. Dark states (illumination states) are treated as negative mood (positive mood), while conductance reaching 29 μS is set as the target memory level [Fig. 11(d)]. When the brain is in a negative mood, the value of conductance triggered by 100 consecutive electrical spikes (amplitude of 1 V, width of 2 ms) reaches 29 μS, which is much larger than that required when the brain is in a positive mood. Through a few seconds of the forgetting process, the conductance values have to be stimulated by 58 (13) consecutive electrical spikes to attain 29 μS again when the brain is in a negative (positive) mood, which means the efficiencies of learning and relearning are controlled by emotions.

More recently, an extensive number of MHP-based optoelectronic synaptic devices, including MHP-based memristors and MHP-based transistors, were employed to fabricate and mimic artificial intelligence visual systems (AIVs) for sensing and processing images [52,62,67,128,184–186]. As an example, Lee et al. reported a dynamic artificial visual adaptation neuron (DAVAN) device with a 3×3 array that mimics the perceptual ability of the human brain using the adaptive ability of the device to the incident light intensity, offering the possibility of constructing an artificial neuromorphic device [59]. A schematic diagram of the human visual system is shown in Fig. 12(a), in which nerve cells in the visual cortex perceive and process the visual information delivered by photoreceptors. As shown in Fig. 12(b), after being repeatedly stimulated, biological nerve cells produce a biological habituation process: the facilitation process (pre) and the inhibition process (post), respectively. To mimic this function, the authors proposed an optoelectronic neuromorphic circuit structured by two neurotransistors and a perovskite photodetector [Fig. 12(c)]. The EPSC of the DAVAN device under the stimulation of repeated light pulses is shown in Fig. 12(d). It can be seen that the dynamic trend of the response current is consistent with the biological habituation process in Fig. 12(b). Figures 12(e)–12(h) illustrate the optical image and dynamic response process of the 3×3 array based on the DAVAN device. Here, the authors successively applied repetitive weak and strong light signal stimuli to the array to simulate the dynamic response EPSC of the array under the photopic vision and scotopic vision conditions. The results show that the device can realize the simulation of the perceptual process of the visual system under different lighting conditions, which has a promising application in the realization of artificial perception systems.

In this review, recent advances in MHP-based optoelectronic synapses with related neuromorphic applications are summarized. Compared with conventional electrical synapses, optoelectronic synaptic devices using synergistic stimulation exhibit unique superiorities such as higher interference immunity and lower energy consumption, which show promise to break the von Neumann bottleneck and build complex neural networks. So far, benefiting from the fascinating properties of MHPs including charge trap, tunable bandgap, and ion migration, MHP-based phototransistors and memristors have been proposed as optoelectronic synaptic devices. The light-stimulated synaptic plasticity (e.g., PPF, STP, STD, LTP, and LTD) of biological synapses and classical Hebbian learning rules was successfully simulated. In addition, several applications of optoelectronic synapses in the fields of neuromorphic computing, higher-order learning, and AVSs were enumerated, which have significant implications for building energy-efficient neuromorphic systems. Although considerable progress has been achieved on the device architectures and applications of MHP-based optoelectronic synapses, there are still various issues worth further exploration.(i)Reducing power consumption. In 2016, the power consumption of Google’s AlphaGo during the confrontation with a human chess player reached 100 kW, far more than the 20 W consumed by the human brain [187–189]. To the best of our knowledge, optoelectronic synapses with low-power characteristics still require operation consumption in pJ or even nJ per synaptic behavior, leading to an urgent and thorny problem of minimizing energy consumption during large-scale array integration. In this case, downscaling the dimensions of the MHP-based optoelectronic synapses while optimizing their structure is regarded as the effective solution proposed so far.(ii)All-optical stimulated synapse. Scientific findings show that human access to external information relies on the biological perception system, and the amount of information transmitted by the human visual system accounts for more than 70% [190,191]. For this reason, light signals for constructing optoelectronic neuromimetic engineering have been introduced into synaptic electronics, which has motivated the emergence of optical-stimulated synapses, optical-assisted synapses, and optical-output synapses. Considering the energy consumption and operational mode, all-optical stimulated synapses are the preferred development direction for future AI compared with complex synergistic stimuli. Because of this, exploiting superior light absorption properties of MHPs for developing all-optical modulated optoelectronic synapses is urgently demanded to facilitate the advancement of neuromorphic devices.(iii)Innovative applications. Until now, the majority of efforts are concentrated on simple simulations of synaptic behaviors, which remain in the initial stage, without a standardized model for evaluating synaptic devices. In terms of MHP materials, synaptic devices possessing a single function responding to optical stimuli alone have already failed to satisfy the demands of AI, and stretchable optoelectronic sensorimotor synapses based on MHPs by their inherently high mechanical flexibility deserve to be exploited as well. In addition, the evolution of these MHP-based optoelectronic synapses remains essential for the development of future neuromorphic systems, such as AIVs, wearable electronics, and human-machine interfaces.