Most of the emerging intelligent applications including autonomous driving, medical imaging, and intelligent manufacturing etc. are related to intelligent vision^[1−3], which can provide the basis for detect and recognize information by simulating biological visual functions, so as to understand the complex external environment^{[4, 5]}. The development of intelligent vision put forward high requirements for the underlying hardware, not only high performances such as high resolution, low delay, and low power consumption etc., but also the flexible functionalities including dynamic response, adaptive learning, and biological compatibility are required^{[6, 7]}. Compared with the traditional vision systems that are composed of multiple discrete parts, neuromorphic vision system owns the advantages of real-time perception, parallel processing and computing^[8]. And it can solve the problems of high delay and high power consumption in the traditional computing architecture^[9]. In this regards, neuromorphic devices that integrating the dynamic visual functions via the ophtoelectric response and the subsequent change in conductance state can provide the underlying fundamental for neuromorphic vision systems^[10]. Therefore, the research and development of neuromorphic devices are of scientific and practical significance.

Biologically, visual adaptation is vital for perceiving, filtering signals, and dynamically processing the information in the varying environment^[11], enabling organisms the capabilities of efficient acquirement, quick processing, and fast reacting, so as to complete specific tasks and avoid potential hazards^[12]. However, achieving the visual adaptabilities in devices are challenging. For example, light signals carry a variety of information, including light wavelength^[13], light intensity, and polarization^{[14, 15]}, etc. Optoelectronic devices should exhibit distinctive responses for various light wavelengths; thus the designation and selectivity of photosensitive functional layers are highly required^[16]. In addition, for the information processing stage, in pursuing the dynamically modulation in the conductance state changes, it is necessary to introduce the multi-terminal regulation and multi-mechanism competition strategies^[17]. More importantly, for mimicking the history-dependent plasticity that enables continuous learning from experience, dynamically analyze the work mechanism and flexibly design the test scheme are required^{[11, 18]}.

The property that optoelectronic devices can perceive external light and the conductive state can be adjusted according to the external stimuli is highly consistent with the changes of synaptic weights in human vision system, including the light sensing in retinal, information transmission through optic nerve, as well as information preprocessing and storage in the visual cortex and brain^[19−21]. Numerous works have reported the two-terminal memristors can exhibit adaptive behaviors through multiple physical mechanisms, including temperature-driven Mott transitions, field-driven defect generation and recombination^[22]. Comparatively, optoelectronic transistors possess the characteristics of multi-terminal modulation and multi-mechanism compatibility, making them great potential in developing neuromorphic vision device^[23−25]. Optoelectronic transistors can generate photogenic excitons by absorbing incident light through photosensitive functional layers^{[31, 32]}. Under a certain electric field, the excitons can dissociate into electrons and holes for charge trapping or channel current, thus generating multi level and reversible conduction states^[27]. When facing the actual scene with continuously varying lighting conditions, the intelligent detection and recognition still facing severe challenges^[33], making the adaptive optoelectronic transistors become one of the main research directions. In recent years, although a variety of adaptive neuromorphic transistors were reported to simulate the adaptive functions, the research is still in the early stage^{[34, 35]}. A comprehensive review and summary of the implementation strategies, research focuses, and challenges in adaptive transistors are crucial for further developing intelligent neuromorphic vision systems.

This review gives the summary of the present research on adaptive optoelectronic transistors (Fig. 1), mainly including the following three aspects. Section 2 discusses the adaptive sensitivity changes in complex light environment, which allows neuromorphic devices to obtain specific signals (light intensity, light wavelength, and polarization, etc.), and directly avoid redundant data at the input terminal. Section 3 summarizes the strategies for achieving adaptive synaptic weight change after stimuli, which amplifies the effective stimuli and extends the response range of light intensity. Section 4 gives the review of the history-dependent plasticity, which enables the threshold sliding or saturation according to the history stimuli, so as to improve the learning or preprocessing capability of the devices. Lastly, Section 5 presents the challenges in developing adaptive neuromorphic devices and the potential opportunities for intelligent neuromorphic vision systems.

In human retina, there are rod and cone cells, mainly playing a role in dark light and bright light conditions, respectively. Moreover, cone cells can be divided into red (R), green (G), and blue (B) cone cells, possessing the RGB color recognition functions. These cells can detect light and convert external light stimuli into electrical signals, forming the beginning of visual information processing. In addition, unlike the trichromatic vision that commonly possessed by human, insects such as bees and butterflies generally own the tetrachromatic vision, laying the foundation for their detecting, processing, and recognizing ultraviolet (UV) information in specific scenes. For the realistic scenarios in intelligent applications, such as robot navigation, automatic driving, and smart medical treatment etc., optical information acquisition is the prerequisite for executing muti tasks. Using optoelectronic neuromorphic devices to adaptively generate the corresponding response according to the light intensity, light wavelength, and specific information of the external light, so as to capture and filter information sensitively and efficiently, is regarded crucial for judging the current situation and making decisions.

The wavelength-dependent response in optoelectronic neuromorphic devices is highly desired for artificial vision systems^[38]. Traditional sensor typically consists three channels for discriminating wavelengths by responding to the RGB incident light independently by using three lateral channel structures with wavelength filters for color component separation, the complex and low space utilization structure limited their opportunities for high-density photodetectors^[39]. In this regards, wavelength-dependent adaptive neuromorphic devices provide solutions for the challenges, the devices allowed a differential response to RGB light in a single device by a wavelength-dependent storage states or the bidirectional optical responses. For example, Park et al.^[36] demonstrated the monolithic integration of vertically stacked RGB quantum dots (QDs) in amorphous indiumgalliumzincoxide (aIGZO) phototransistor arrays (Fig. 2(a)). The direct photochemical patterning process enabled the pixelation of the vertically stacked QDs without using a photoresist, avoiding the chemical contamination, enhancing resolution, and decreasing fabrication complexity (Fig. 2(b)). Based on the vertically stacked device, a highresolution photodetector array with areal device density of 5500 devices cm^–2 (Fig. 2(c)) than that of solutionprocessed flexible photodetectors (200–1600 devices cm^–2)^[40].

For developing the adaptive optoelectronic transistors for RGB discriminating, the elaborate calibration of materials for sensing RGB are of great importance. By employing a size-mixed QD layer of ratio-controlled QDs with three different optical bandgaps that can sense R, G, and B as the photo absorbing layer. Park et al.^[13] also developed the heterogeneous phototransistors for distinct wavelength recognition (Fig. 2(d)). The precise design of QD mixing ratio allowed the quantitative tuning of wavelength-specific responses, also, a simplified device and system fabrication process were achieved (Fig. 2(e)). Further, the wavelength-dependent synaptic plasticity were also obtained via the obviously distinguished charge storage capability (Figs. 2(f) and 2(g)). For another example, benefiting from the broad selection and wide tunable energy band structures of van der Waal semiconductors, Lei et al.^[41] designed a sensor with excellent photoresponse by stacking CIS, InSe, and GaS as photosensitive layers to sensing R, G, and B, respectively (Figs. 2(h)–2(j)). Through the spectral response and the linear dependence of photocurrent on incident light power in each sensing channel, the white-balance calibration was achieved and the implemented device can effectively recognize different light wavelengths (Figs. 2(k) and 2(l)). The authors also confirmed the feasibility and effectiveness of the design strategy by the device functionality tests, including the spatial resolving of light intensity distribution and color recognition.

The devices that can perform bidirectional optical response under RGB light provided another strategy for wavelength dependent adaptive devices. Huang et al.^[26] developed a heterojunction phototransistor with two-dimensional (2D) CsPbBr₃ nanoplates/poly(3-hexylthiophene-2,5-diyl) (P3HT) as the photoactive layer and tunnel layer. The absorption range of CsPbBr₃ nanoplates was complementary to that of P3HT, providing the basis for the wavelength-dependent plasticity in the transistor. By changing the wavelength (405 or 620 nm) of the incident light, excitons were generated in CsPbBr₃ or P3HT photoactive materials, then, the wavelength-dependent positive and negative photoconductivity (PPC and NPC) in the devices can be achieved with the aid of gate voltage modulation. Based on the unique synaptic plasticity, the authors simulated the Drosophila’s evasive behavior towards food under red and blue light, demonstrating an opportunity for the device simulation of biological visual functions (Figs. 2(m) and 2(n)). The similar strategy for distinguish different light wavelengths by the PPC and NPC effect can also be seen in the research reported by Jiang et al.^[37]. They fabricated a spontaneous chromatic adaptation in the two-terminal bilayer PtSe₂ transistor. In contrast to the gate-controlled bipolar photoconductivity, the wavelength-dependent PPC and NPC were obtained under the irradiation of 650 light and 450/532 nm light, respectively (Fig. 2(o)). The phenomena were originated from the photocontrolled physical adsorption and desorption of oxygen (O₂) molecules on the bilayer PtSe₂ surface. Moreover, red–blue and red–green antagonistic receptive fields are demonstrated (Fig. 2(p)), which enabled the device the unique capability for distinguishing the warm targets from cool backgrounds.

In addition to visible light, the invisible light often carries special information in the external environment. Integrating the functions of invisible light detection in optoelectronic transistors would make the artificial vision system beyond the limit of human vision. On the other hand, UV light with the wavelength of 320–400 nm is harmful to human retina^[42−44]. In view of this, exploring devices that can adaptively detect the specific UV light is necessary for early warning of ultraviolet radiation. However, in the actual ultraviolet detection, it is crucial to detect the commonly weak UV light and overcome the interference of other light wavelengths at the same time^[45]. In this regards, Ji et al.^[17] selected the asymmetricmolecular structured 2-hexylthieno[4,5-b][1] benzothieno[3,2-b][1] benzothiophene (BTBTT6-syn) organic small molecule semiconductor with unique UV absorption as the active layer in optoelectronic synaptic transistor (Figs. 3(a) and 3(b)). Ultraweak UV light (370 nm) with light intensity as low as 31 nW∙cm⁻² was adaptively detected by the synaptic transistors (Fig. 3(c)), while for the R (650 nm), G (520 nm), and B (450 nm) light irradiation, the postsynaptic current (PSC) and weight update values were much lower, indicating the high UV selective response and making it possible for the device to extract the weak UV signal from the interfered RGB noise (Fig. 3(d)).

The demand for ultraviolet detection brought great challenges to the research of wide bandgap materials that can respond to UV light, especially the selective response in solar blind light (200–280 nm short-wave UV), which is of great significance for aerospace exploration, atmospheric monitoring, and health care etc.^[46], The current strategies for detecting solar blind light rely on the employment of expensive optical filters or wide bandgap inorganic semiconductors. Xu et al.^[47] reported the optoelectronic synaptic transistors by using 2D wide bandgap semiconductor Ca₂Nb₃O₁₀ with high responsivity and selectivity to deep UV rays (Fig. 3(e)). The synaptic devices can distinguish UV light with different wavelengths (365, 310, and 254 nm), and were extremely sensitive to UV light (254 nm) at a low intensity of 70 nW∙cm⁻².

The designability of organic materials offered another possibility for the detecting solar blind light. Wang et al.^[48] reported a solar-blind UV monitoring by utilizing pentacene and N,N’Ditridecylperylene-3,4,9,10-tetracarboxylic diimide (PTCDI-C₁₃) as the tunnel layer in the transistors. These organic transistors can adaptively measure the solar-blind UV light (254 nm) in a nonvolatile and rewritable manner, but cannot respond to the UV light with 365 nm or visible light. The authors proposed a UV-induced interfacial excitation mechanism to interpret the device features. In the mechanism, the selective and storable UV response was ascribed to the UV-induced charge trapping in the poly (2-vinyl naphthalene) (PVN) layer at a small gate field, the UV light must possess sufficient photon energy to enable the interfacial charge excitation and subsequent charge hopping into the PVN bulk.

Aside from UV detection, infrared detection also played an important role in various fields, such as thermal imaging, smart night vision, and medical diagnostics, etc. For efficient absorbing mid-infrared light and accurate encoding based on random the mid-infrared light sampling, Wang et al.^[49] constructed the optoelectronic transistors with b-AsP/MoTe₂ 2D heterostructure. In the device, b-AsP with a narrow bandgap (0.15 eV), a high mid-infrared light absorption efficiency (10%), and a high hole mobility (145 cm²∙V^–1∙s^–1) was used as the mid-infrared photosensitive layer, and MoTe₂ was used as the mid-infrared sensitizer, the heterostructure promoted the charge separation and transport. For another example, Ling et al.^[50] designed the optoelectronic transistors with three-dimensional porous pentance/PM6:Y7-BO, taking advantages of the efficient charge separation between PM6 and Y7-BO, the remarkable light absorption ability of PM6, and the ordered crystalline structure of Y7-BO, the light absorption efficiency was enhanced by 2.5-fold. Meanwhile, the authors also demonstrated the bidirectional response capabilities of non-volatile PPC and volatile NPC to visible and near-infrared light, which enabling the image edge detection in varying lighting conditions (with precision above 85%).

Aside from the light wavelengths information, the detection of polarized light also attracted the attention. Polarized light is ubiquitous in nature. It can reveal the characteristics of the object, such as material composition, surface shape, and dielectric constant^[51]. Moreover, real-time navigation can be achieved via monitoring the sun-position-related celestial polarization cues^[52]. With various application scenarios and prospects, the advanced polarization sensitive visual adaptation in optoelectronic devices was still lacking and rather challenging^{[53, 54]}. In this regards, Selecting proper functional materials that can sense polarized light in optoelectronic transistors is of vital importance^[15].

Compared with the polarized light sensitive metamaterials^[55], anisotropic^[56], and polymer crystals^[57], 2D polarimetric devices with atomic-level anisotropy own the potential of high-density integration and scalability^[58]. By introducing the porous metal–organic-framework (MOF) material (UiO-66-NH2, U6N) into anisotropic-ReS₂-based phototransistor, Jiang et al.^[14] demonstrated the synaptic device with polarization-sensitive visual adaptation. The device exhibited efficient polarized light perception due to the strong in-plane anisotropic crystal structure and direct bandgap which can be ascribed to the distorted 1T phase and weak interlayer coupling in 2D-ReS₂^[59−61]. Moreover, the efficient functional group of amino group (–NH₂) in U6N can promote the charge-transfer interactions in U6N^{[53, 54]} and the porous structured U6N also can facilitate the photogenerated charge separation^[53]. The authors also validated visual adaptation of the green light (552 nm) or the polarized light (860 nm, 50 mW∙cm⁻²) with the aid of gate voltage (10 V), and they ascribed the adaptation to the charge trapping and de-trapping mechanism at ReS₂/U6N interface. For another example, Xu et al.^[15] developed an optically-controlled polarimetry memtransistor based on a van der Waals heterostructure (ReS₂/GeSe₂) with anisotropic structures (Fig. 3(f)). The devices exhibited PPC and NPC properties under 808 and 405 nm, respectively (Fig. 3(g)). The polarimetric detection property enables the device to identify the celestial polarizations for real-time navigation (Figs. 3(h) and 3(i)). Moreover, the authors also demonstrated that with anti-glare function, the training epochs during the pattern recognition was significantly reduced, and the training energy consumption was reduced to 11.4% (Figs. 3(j) and 3(k)).

Aside from perceiving external light information, biological vision system can also complete information preprocessing, recognition, and memory functions. In fact, biological nervous system consists of a complex central nervous system and peripheral nervous system, the vast network formed by large amount of neurons and synapses makes it one of the most efficient computing and storage systems. The realization of these functions depends on the dynamic changes in the connection strength between synapses dynamic changes in the connection strength between synapses (i.e., synaptic plasticity), which mainly includes short-term and long-term plasticity. Short-term synaptic plasticity mainly includes facilitation, depression, and potentiation, while long-term synaptic plasticity is mainly manifested in the form of long-term potentiation (LTP) and long-term depression (LTD). LTP and LTD were recognized as the biological basis of learning and memory activities. Furthermore, the adaptive biological vision function that the plasticity can dynamically change with the varying external environment, plays an important role in the dynamic preprocessing and recognition processes. In this regards, much efforts were paid to mimicking such functions in optoelectronic devices, and the functions of dynamic short-term potentiation (STP) and LTP, photopic and scotopic adaptation, and active adaptation were achieved.

Generally, STP and LTP were responsible for preprocessing and recognition processes, respectively. For their realization in optoelectronic synaptic transistor, the key point is the decay rate of conduction states. STP was considered as the synaptic plasticity for perceiving and processing external information, the corresponding time scale is basically a few milliseconds to minutes. While LTP is regarded as the primary mechanism for learning and memory due to its longer time scale, ranging from tens of minutes to hours^[62−65]. It is widely reported that the transformation from STP to LTP in optoelectronic synaptic transistor can be obtained by increasing the stimuli dose, such as increasing illumination durations, intensities, and frequencies^{[66, 67]}. However, when processing the extensive unstructured data in vision system, the rapid and repeated switching between various functions (preprocessing, recognition, and memory functions) is necessary. Moreover, under the complex lighting scenes, visual selective attention is also essential for filtering the irrelevant information and enabling humans to manage the most effective stimuli through processing the salient regions and suppressing the non-salient regions^[68−73]. In this regards, hardware implementation of the adaptive functions will provide solutions for developing advanced artificial vision system^[74−77].

Facing the challenges and application prospect, Huang et al.^[27] designed a reconfigurable neuromorphic transistor (Fig. 4(a)), in which, an air-stable Zr-CsPbI₃ perovskite nanocrystal was prepared and introduced. When Zr-CsPbI₃ PNCs absorbed the 405 nm light with high photon energy, the photoinduced electrons were captured by the deep trap state and introducing non-volatile property, while for the 650 nm with low photon energy, the electrons generated could not reach the deep trap state in the conduction band to recombine with the hole, and resulting the volatile property, Thus, benefiting from the differential electron trapping ability, STP and LTP were implemented separately under 650 and 405 nm incident light (Figs. 4(b)–4(d)). Moreover, STP and LTP could be switched in a rapid and repeated way by changing the wavelength of optical stimuli. The monolithic integration of wavelength-dependent reconfigurable STP and LTP functions enabled the realization of signal filtering, blue target in images was then extracted in the preprocessing part, thus effectively improving the recognition accuracy (Figs. 4(e)–4(j)). Another strategy for integrating different decay rate in a single device is reported by Chen et al.^[78]. They introduced a programmable ferroelectric functional layer in the poly [2,5-bis (alkyl) pyrrolo-[3,4-c]pyrrole-1,4(2H,5H)-dione-alt-5,50-di(thiophen-2-yl)-2,20-(E)-2-(2-thiophen-2yl)vinyl)thiophene] (PDVT-10) optoelectronic transistor. The promising ferroelectric material poly(vinylidene fluoride-trifluoroethylene) (P(VDF-TrFE)) with tunable remnant polarization that can modulate the charge accumulation and depletion was utilized (Fig. 4(k)). Under the positive voltage, the upward polarization domain gradually promoted the photogenerated holes tunnel through the energy band and impeded the recombination of photogenerated carriers, leading to LTP. While for the negative polarization, the polarized electric field blocked the migration of the photogenerated holes, and the holes in semiconductor dissipated rapidly, thereby manifesting the short-term relaxation characteristics. Memory strength in the transistors can be modulated from 9.1% to 47.1% without peripheral storage unit (Fig. 4(l)). The authors also demonstrated the selectively recorded and suppressed UV light information by simply vary the polarization direction of P(VDF-TrFE) layer. Moreover, based on the unique synaptic plasticity, the selective attention function was mimicked, and the corresponding sensory network can exhibit a classification accuracy of 95.7%.

The rapid development of bionic artificial vision system promoted the application in the emerging intelligent products, such as automatic driving, intelligent robots, and deep-space exploration. In these scenarios, high resolution, high detection speed, and the ability for detecting a wide range of lighting intensity were required^{[8, 80−82]}. In human vision, photopic and scotopic adaptation plays an important role in coping with the suddenly changed ambient light intensity. Specially, when suddenly entering a dim environment from a bright ambience, a person will initially be unable to see anything, and then can gradually see the surrounding objects as the visual system adapts to the dark light environment. This phenomenon is called the scotopic adaptation. Correspondingly, when suddenly entering a bright environment from a dim background, a person instinctively feel dazzled and can gradually see the objects after a period of time for photopic adaptation^[83] (Fig. 5(a)). In fact, rod cells and cone cells in human retina are responsible for capturing visual information with dim and bright light, respectively (Figs. 5(b) and 5(c)). Both of the cells possess a limited detection range (40 dB) for light illumination, and their combination and the adaption functions enabled a wide response range of light intensity (over 160 dB)^[1].

Inspired by the photopic and scotopic adaptation, Chai et al.^[1] designed a vision sensor array based on the bottom-gated bilayer MoS₂ optoelectronic transistors, in which, charge trap states were introduced into MoS₂ surface, enabling the long-term memory property. Moreover, the authors demonstrated the time-varying excitation and inhibition characteristics under negative and positive gate voltage, respectively. And the localized trap states (S vacancies) in MoS₂ that distributed over a broad energy range in the bandgap and serving as ambipolar trap states under different gate voltages were considered as the key point for the dynamic modulation. In this regards, scotopic and photopic adaption were mimicked in the device, and an effective perception range of 199 dB was achieved (Fig. 5(d)). Based on the basic characteristics, an 8 × 8 device array was fabricated to perceive "8", through which, the enhancement in image contrast was achieved under various light conditions, thus improving the recognition accuracy. Aside from this, a number of devices were used to mimic the photopic and scotopic adaptation^[84−86]. For example, Sun et al.^[87] presented a phototransistor comprising a sandwich structured graphene/PbS QDs/graphene heterojunction, in which, the PbS QDs film serving as the near-infrared sensing layer. The charge trapping and de-trapping processes controlled by the gate voltage allowed the time-varying excitation and inhibition characteristics, thus achieving photopic and scotopic adaptation under near-infrared (1064 nm) light. Moreover, by embedding the discrete InP QDs within the InSnZnO (ITZO) thin film, Cao et al.^[79] created an optoelectronic transistor with InP QDs/ITZO serving as the channel layer. Benefiting from the energy band alignment and gate control features of the transistor, the authors demonstrated an enhanced visible light responsivity, a linear response to the visible light and the scotopic and photopic adaptation by the aid of gate voltage.

The research on scotopic and photopic adaptation has been extended to the polarized light detection and the wearable electronics. In the research by Liu et al.^[88], the wafer-scale arrayed artificial optoelectronic transistors based on a pentacene/chiral nanoclusters heterostructure were created (Figs. 6(a) and 6(b)). In the transistor, nanoclusters with light assisted and tunable Fermi level were employed to control the charge trapping and de-trapping processes, enabling the spectral-dependent visual adaptation. And benefiting from the chirality of nanoclusters, the device can perceive the circular polarization information (Fig. 6(c)). The integration of multi-task features provided an opportunity for the structurally simplified all-in-one artificial vision systems (Fig. 6(d)). For developing the advanced wearable and adaptive optoelectronic devices^[89−92], Guo et al.^[93] designed an intrinsically stretchable (poly(3,6-di(2-thien-5-yl)−2,5-di(2-octyldodecyl)-pyrrolo[3,4-c]pyrrole-1,4-dione)thieno[3,2-b]thiophene) (DPP-DTT)/SEBS transistor (Fig. 6(e)), in which, the viscoelastic photosensitive films composed of CsPbBr₃ quantum dots and SEBS elastomer ensured the broad-wavelength light detection (Figs. 6(f) and 6(g)). Through sequent thermal lamination and transfer procedures, defect-tunable interfaces and ideal contact interface between the intrinsically stretchable multi functional layers were achieved, enabling the trichromatic photoadaptation and high biaxial stretchability (up to 100%) in the devices. Moreover, the authors also emulated the vision-adaptive imaging by fabricating a 5 × 5 transistor array, and this approach accelerated the implementation of the advanced neuromorphic vision systems.

In addition to the photopic and scotopic adaptation enabled by the cone and rod cells, visual adaptative functions also perform an autonomic response to stimuli, providing sensory system the ability to adjust its behavior under the constantly changing light conditions^[94]. Mimicking the active adaptation in optoelectronic device that can provide a direct response to stimuli is important for developing the energy efficient and space saving bionic vision systems^{[95, 96]}. However, such an active photoadaption device should possess both light intensity-dependent transient response and dynamic adaptation without changing other stimuli, and both excitation and inhibition processes should be exhibited in a single device, which were difficult due to their needs of two contrary charge carrier modulation in the same conductive channel.

Facing the challenges, the strategy of integrating multiple mechanisms in a single device was reported and widely developed^[97]. Di et al.^[28] showed an optoelectronic transistor with two complementary heterojunctions (poly{2,2′[(2,5-bis(2-hexyldecyl)-3,6-dioxo-2,3,5,6-tetrahydropyrrolo[3,4-c]pyrrole-1,4-diyl) dithiophene]-5,5′-diyl-alt-thiophen-2,5-diyl} (PDPP3T):[6,6]-phenyl-C₆₁-butyric acid methyl ester (PCBM) and P3HT:PCBM) (Fig. 7(a)). The authors demonstrated that a photovoltaic effect induced excitation and the charge trapping dominated inhibition in the transistor, the coupling of the two effects enabled the active photoadaptation behavior that the decay rate of PSCs dependent on the light intensity was achieved without changing other stimuli (Fig. 7(b)). Specially, for the upper PDPP3T:PCBM heterojunction, photogenerated holes were accumulated at PDPP3T:PCBM/poly(vinyl-cinnamate) (PVCN) interface, facilitating the charge transport, while for the P3HT:PCBM heterojunction, the photogenerated electrons were driven to accumulate at P3HT:PCBM/polyvinyl alcohol (PVA) interface and eventually trapped by PVA, causing the shielding effect of the gating field (Fig. 7(c)). Thus, the authors concluded that photocarrier generation in the two heterojunctions, dynamic charge trapping in the composite dielectric layer, and charge transport in the conductive channel were responsible for the active adaptation. Further, a flexible 3 × 3 array (2 × 2 cm²) was fabricated, the device array can produce the projecting T-shaped images before and after the expose of strong light, demonstrating the advantages of the adaptive transistor in mimicking the human visual adaptation.

Aside from introducing multi-functional layers and avoiding the difficulty in material selection, Ling et al.^[98] reported the nano-patterning engineering strategy for the light intensity-dependent decay rate in a nanoporous poly(N-vinylcarbazole) (PVK) templated pentacene transistor (Fig. 7(d)). It was demonstrated that the localized electric field generated by nanopouous structure can improve the light absorption in both PVK dielectric and pentacene channel layers (Figs. 7(e) and 7(f)), and facilitate the charge injection process, thus benefiting the efficient charge trapping effect. The limited carrier transport and hindered electron–hole recombination in nanoporous devices resulting in the smaller decay rate under the low light intensity. While for the higher light intensity, benefiting from the improve generation and separation of charge carriers, as well as the charge injection efficiency in nanoporous devices, the charge de-trapping process were promoted due to the increased space charge shielding effect, resulting in a higher decay rate. Thus, the adaptive decay rate was realized via the light intensity-dependent dynamical charge trapping and de-trapping processes (Figs. 7(g) and 7(h)). Moreover, by constructing a biomimetic eye with visual selective attention, eye rotation, and eye closing functions were realized. And the non-ideal image with 0.8% contrast was precisely preprocessed, achieving a high recognition accuracy of 95.4% (Fig. 7(i)).

When performing preprocessing and memory functions, synaptic plasticity is influenced by muti active factors, such as the spike intervals^[99], presynaptic and postsynaptic firing rates^[100], and temperature, etc.^[101]. The previous stimuli as the history factors can also impact on the synaptic efficacy for the subsequent stimuli^[102], enabling humans the ability to continually learning from experiences. Implementing such complex learning capability in optoelectronic devices is considered a key step towards achieving neural networks with fast, high-precision computing property. For example, the neuromorphic systems typically implement the spike timing dependent plasticity (STDP) that maps synaptic weight changes as a function of the time between presynaptic and postsynaptic spikes^{[103, 104]}, this simple time-based model is very convenient due to the low-power operations within a defined domain. In addition, history-dependent plasticity enables brain the ability of continually evolving and learning from the previous experiences, thus changing the responses to the subsequent stimuli. Mimicking the history-dependent plasticity, such as the synaptic metaplasticity and synaptic saturation in optoelectronic transistors will provide feasibility for enhancing the learning and computing capabilities in neural networks.

Synaptic plasticity in biological neural system is not only activity-dependent, but also sensitive to the previous activities (i.e., history-dependent). Metaplasticity is a phenomenon that defined as the plasticity of synaptic plasticity, in this phenomenon, the threshold (θ⁺) for LTP and LTD can be modulated by prior activity^[105]. In biological system, the binding of information stored at neighboring synapses and branchlets can be achieved by lowering the θ⁺, and the raise of θ⁺ can avoid the synapse from over-activation^{[102, 106]}. The direction or degree of synaptic plasticity is determined by the prior input signals, thus realizing the second-order adaptation function, which is crucial to accelerate learning rate and improve learning efficiency. For the synaptic plasticity, LTP and LTD are deemed to be the fundamental of learning and memory for information processing and storage^{[107, 108]}, while metaplasticity is considered crucial for improving memory and learning capability^[109]. Metaplasticity inspired the stable and efficient learning, and was introduced into neuromorphic computing system to overcome the existing issues in the algorithm, such as reducing the catastrophic forgetting^[110], improving the learning rate and recognition accuracy in image classification^{[111, 112]}.

For device implementation, Fig. 8(a) showed the schematic drawing of the vital features in metaplasticity: 1) the long-term plasticity^[113−115]; 2) the threshold sliding of LTP and LTD; 3) the non-monotonic behavior for enhanced depression effect (EDE) region^[113]; 4) the small threshold (θ^–); 5) the multiplicative relationship between presynaptic and postsynaptic neuron activities^{[116, 117]}. The most prominent theoretical or computational model that relates directly to metaplasticity is the Bienenstock, Cooper, and Munro (BCM) rule, in which, synaptic weight change is frequency-dependent. By using the competition between the forgetting effect and the stimulus effect, the researchers have reported the realization of the frequency-dependent threshold sliding phenomenon in the second-order memristor^{[118, 119]}. In optoelectronic transistor, various researches have also reported the BCM phenomena^[120−123]. For example, Wan et al.^[120] proposed an amorphous indium gallium zinc oxide (a-IGZO)-based optoelectronic transistor that can mimic BCM learning rule. In the transistor, SiO₂ electrolyte with a large electric-double-layer capacitance (0.33 μF∙cm^–2) was used as the gate dielectric. By setting various frequencies as the different active histories, the authors demonstrated the history-dependent synaptic plasticity and the frequency-dependent threshold sliding in BCM, the results were attributed to the competition between the generation of photoelectrons under light and spontaneous neutralization of photoelectrons after light. For another example, in the research work reported by Pi et al.^[123], a single optoelectronic transistor based on the hybrid structured silicon nanocrystals and poly(3-hexylthiophene) (P3HT) was proposed. Setting the electrical spike with the amplitude of +15, 0, and –15 V as the different active history, the authors realized the threshold sliding in metaplasticity. Moreover, based on the phenomenon that the synaptic weight change was dependent on the spike frequency, and assuming the mutation of an image with a frequency higher (lower) than the threshold was enhanced (ignored). The authors also simulated the high-pass filtering, through which, the details (edges) can be detected with lower (higher) threshold conditions, demonstrating the potential applications for realtime image processing ( Figs. 8(b) and 8(c))^[124].

However, the above implementation of metaplasticity and BCM rules in memristor and transistor were focused on the history-dependent plasticity and the threshold sliding effect with the monotonic dependence on the spike rate, the EDE in the original metaplasticity in neurobiology was not well mimicked^{[119−123, 125]}. Recently, based on the numerical simulation proposed by Pfister et al.^{[126, 127]}, the generalized BCM rule with EDE was realized in the Pt/WO_3−x/W second-order memristor by introducing the triplet-STDP instead of rate-dependent spikes (Figs. 8(d) and 8(e))^[117], and the rate-based orientation selectivity was further demonstrated in a simulated feedforward memristive network. Following, by designing the van der Waals heterostructure (MoS₂/WSe₂) memtransistor^[128], the memristor with Au/P3HT/CsPbBr₂I/ITO structure^[129], and ITO/SnO₂: [6,6]phenyl-C61-butyric acid (PCBA)/methylammonium lead iodide (CH₃NH₃PbI₃)/poly(3-hexylthiophene2,5-diyl) (P3HT)/MoO₃/Ag memristor^[116], the generalized BCM rule with EDE were further demonstrated in various devices (Figs. 8(f) and 8(g)). Moreover, based on the devices, the rate-based orientation selectivity in a neuron network, the bioinspired striate cortex with binocularity, and the orientation selectivity and simulation of binocular orientation-selective neural networks were developed, respectively (Figs. 9(a) and 9(b))^{[116, 129]}.

In addition to frequency-dependent BCM rule, evidence for voltage-dependent metaplasticity was reported in hippocampus^[107]. However, the direct and complete simulation of metaplasticity still remained a difficulty^[125]. Facing this challenge, Huang et al.^[29] designed a dually adaptable optoelectronic transistor with p–n pentacene/PTCDI-C₁₃ heterojunction. By increasing the amplitude of the applied gate pulse, the efficient charge trapping effect and the switching from p-type to n-type channels in heterojunction were obtained (Fig. 9(c)), enabling the unique unipolar spike voltage-dependent plasticity (U-SVDP) in the device^{[130, 131]}. History-dependent property was also obtained benefiting from the good memory characteristics (Figs. 9(d) and 9(e)). Based on these, metaplasticity was then achieved (Figs. 9(f) and 9(g)). In addition, the photoresponse property caused by the built-in potential in heterojunction provided the basis for negative photoconductivity and enabled the light intensity-dependent threshold sliding in the transistor. Thus, adaptive binary thresholding in local subimages was used for image contrast enhancement. Further, by constructing a convolutional neural network architecture with U-SVDP as the convolution kernel and the threshold sliding for adjusting the threshold of U-SVDP during the backpropagation process, the dual-adaptive heterojunction synaptic transistor enabled the precise preprocess of the low contrast (0.4%) image, the convergence rate was significantly improved by 5 times, and the recognition accuracy improved to 93.8% was also obtained (Figs. 9(h)–9(k)). These results demonstrated the device as a potential for future visuomorphic computing under harsh lighting conditions.

According to the Hebbian learning rule, synaptic activity can increase the synaptic strength, leading to more activity and further modification. However, the unlimited excitation or inhibition under continuous stimuli may result in the uncontrolled change in the synaptic strengths. A synaptic saturation, i.e., a limited value in synaptic weight, as an inherent constraint to modify Hebbian learning rule was needed. When reaching the saturation, no LTP or LTD would be further induced, thus preventing further changes in the synaptic weight^[132]. In optoelectronic transistor, the limited value can be obtained via the saturation of charge trapping during the continuous stimuli, and the amount of variation in synaptic weight relied on an effective stimuli dose, which was determined by the pulse amplitude, pulse width, and pulse number etc. Ling et al.^[133] reported an optoelectronic transistor with strong UV absorption material Poly[(9,9-dioctylfluorenyl-2,7-diyl)-co- (4,4′-(N(4-sec-butylphenyl) diphenylamine)] (TFB) as the photosensitive layer (Fig. 10(a)). With increasing the pulse number of the stimuli, the device exhibited LTP and STP successively (Fig. 10(b)). The electrical conductance of the device under UV radiation was used to simulate the object perception before and after the retinal damage. Moreover, by constructing an electronic eye system, the pupil constriction and eye closing behaviors were presented, which expanded a concept for the future development in visuomorphic computing.

In addition, the charge trapping effect can also form a space charge region, inducing a reverse built-in electric field and shielding the gate voltage. Therefore, a decrease in the PSC might be obtained after the synaptic saturation. For example, Chen et al.^[30] designed a Indacenodithiophene-benzothiadiazole (IDTBT) synaptic transistor (Fig. 10(c)), in which, polyvinyl alcohol (PVA) was employed as the charge-trapping layer, and an offluorinated iso-indigo [7,6-g] iso-indigo (P0FDIID)/(poly{[N,N-bis(2-octyldodecyl)-naphthalene-1,4,5,8bis(dicarboximi-e)-2,6-diyl]-alt-5,5’-(2,2’-bithiophene)]} (N2200) heterostructure that can absorb broad-spectrum light was introduced and served as the photoinduced shielding layer^{[134, 135]}. By coupling the space charge potential and the gate potential, excitation and inhibition occurred successively under the irradiate of a constant light intensity. Therefore, the device can be transformed from a linear model to a nonlinear model, enabling an adaptive tone mapping for static information and achieving a high recognition accuracy (over 90%) for dynamic information. Aside from this device, a similar phenomenon was also reported in the synaptic transistors based on an organic single crystal phototransistors. By manipulating the charge dynamics of the trapping centers of organic crystal-electret vertical stacks, the device firstly exhibited an excitation behavior, and with increasing the pulse number of the stimuli, an inhibition synaptic plasticity was observed after a saturated conduction state. Based on this, Hu et al.^[75] developed a neuromorphic visual system that can successfully distinguish the overexposed image with high recognition accuracy (98.2%) (Figs. 10(d) and 10(e)).

Neuromorphic transistors provide the underlying devices for developing high-performance and multifunctional neuromorphic vision systems, due to their structural and functional similarities with biological vision systems. Recently, facing the continuously varying lighting conditions, adaptive neuromorphic transistors have attracted much attention and become one of the main research directions. In this review, we described the biological adaptive functions and we summarized the representative strategies for achieving these adaptabilities in optoelectronic transistors. The adaptive functions were divided into three classes, including the adaptation for detecting information, adaptive synaptic weight change, and history-dependent plasticity. For each of the adaptive functions, we also pointed out the key focuses in the strategies. Further, the corresponding applications of the achieved adaptive optoelectronic transistors are illustrated.

In future research, several factors require further consideration: 1) Photosensitive materials with wide absorption spectrum or specific light response that are favorable for detecting information with wavelength-dependent adaptation need to be further designed and synthesized. 2) The flexible construction of device structures with multiple mechanisms that can perform differential charge transfer, ion migration, and built-in electric field changes should be considered to achieve the flexible synaptic weight change after stimuli and the history-dependent plasticity. Introducing multi-mechanisms for the collaborative and competition in constructing novel devices might be an efficient way. 3) To meet the application requirements in specific scenarios, such as the human-machine interaction, future studies on neuromorphic transistor also should pay attention to developing devices with flexibility, stretchability, and biocompatibility. In addition, the integration of multiple adaptive functions into one single device would be an efficient way for the application of neuromorphic vision system. 4) Furthermore, current reported adaptive optoelectronic transistors mainly focused on the implementation of adaptability in one single device, large-scale integration should be investigated to further advance the development of intelligent applications. The uniformity between devices, carefully design of the circuit layout, and the heat dissipation issue might be the key matters and difficulties.

Although some challenges remain in developing the adaptive optoelectronic transistor for neuromorphic vision system, we believe that through the combined research efforts in material synthesize and device mechanism analysis, the practical application of neuromorphic vision systems based on adaptive neuromorphic transistors will be realized in the near future.