Advances in Multi‐Terminal Transistors as Reconfigurable Interconnections for Neuromorphic Sensing and Processing

Inspired by the brain, neuromorphic computing seeks to improve data‐centric applications in terms of power consumption and device footprint. The processing efficiency can be attributed to reconfigurable connections coupled with complex dynamics and efficient data sensing and signal processing. With field‐effect tunability, the three‐terminal thin‐film transistor is initially explored as a switch. Recently, transistors with added gates and modes have been fashioned as multi‐terminal devices capable of interconnecting complex networks and transducing multiple inputs. Here, the advances in both the semiconducting channel material and the configurational design of multi‐terminal devices are reviewed. The use of strong coupling and dynamical material properties in novel ion‐based capacitors, mixed‐ionic‐electronic semiconductors, low‐dimensional confined systems, stimuli‐responsive materials and redox‐active semiconductors has enabled new neuromorphic functionalities. These include improving reconfigurability using the regulatory ability of homeostasis, dendritic integration of sensory signals and the synaptic competition of resources. The transistor devices are also highly relevant in circuits that improve the efficiency of sensory transduction and signal processing; stimuli‐responsive semiconductors that adapt to the sensory inputs can be employed in the transistors and highly efficient sensorimotor connections between sensors and actuators can be emulated. These highly efficient, low‐complexity analog devices are crucial design elements for next‐generation neuromorphic systems.


Introduction
Digital electronics have enabled the development of artificial intelligence (AI) and machine learning (ML), imbuing machines with the cognitive capabilities to perform tasks natural to humans.Coupled with continual improvements both in transistor scaling (hardware) and algorithms (software), AI has achieved tremendous impact in various technological fields ranging from Large Language Models, [1] text-to-image generation [2] to autonomous vehicles [3] and humanoid robots. [4]Currently, the capacity originates from computational models trained on huge datasets, finetuning billions of extensively connected parameters known as weights.Optimizing these weights and computing the output of these models would require either tremendously high-speed sequential or massively parallel processing.However, sequential processing based on the von-Neumann architecture (where memory and processor units are separated) is extremely inefficient due to the constant shuttling of data between the two units.This restriction in configuration, generally known as the von-Neumann bottleneck, impedes throughput and drastically increases energy usage.This bottleneck is what neuromorphic engineers target by emulating the building blocks of biological neural networks themselves -the neurons and synapses which co-locate memory and processing.The biological processing network vastly differs from the conventionally used von-Neumann architecture.The human brain comprises an extensively interconnected network of 1012 neurons bridged by 10 15 synapses, resulting in high computational efficiency and low power consumption (≈20 W). [5,6] These impressive capabilities originate from the human brain's unique operational architecture based on synaptic activity -allowing both information processing and storage to occur simultaneously.
Learning in the brain occurs primarily via plasticity -the strength of synaptic connections between neurons.The cascading of ion and neurotransmitter signaling pathways lead to the firing of action potentials, which modulates the strength of synaptic connections or synaptic weights. [7,8]Temporary and permanent changes to this synaptic weight underlies the memory formation process and forms the foundation of all learning rules in the human brain. [9]By incorporating the aforementioned bio-inspired reconfigurable connections, neuromorphic computing is expected to enable profoundly parallel computing and highly efficient processing for artificial intelligence.Such approaches to low power and efficient computation also find strong applications in robotics, intelligent wearables, and ubiquitous sensing.As learning in the brain is governed by chemical synaptic activity, mimicking synaptic signal modulation and transmission through electronic circuitry is key. [10]While artificial synapses based on the Differential Pair Integrator circuit [11] have been fabricated on Complementary Metal-Oxide-Semiconductor (CMOS) technologies, it is challenging to miniaturize the design comprising of several transistors and capacitors.The emergence of nanoscale memristors with intrinsic dynamics resembling biological synapses provides a facile yet effective way to fabricate neuromorphic hardware.With operational history/memory, atomic-level scalability, non-volatility, low energy consumption, and ultrafast switching speeds, nanoscale memristors have been extensively investigated for brain-inspired architectures in recent years. [12]econfigurability in connections or synaptic plasticity has extensively been demonstrated in these two-terminal devices. [12,13]o emulate synaptic plasticity, a variable channel conductance with reliable analog properties such as non-abrupt switching property, continuously distributed resistance states, and repeatable behavior, is needed. [14,15],17] These have been extensively covered in several review articles [18][19][20] and are not within the scope of this review.However, such twoterminal devices face issues with weight precision due to the stochastic nature of filamentary growth.The restriction to electrical stimulus across a single direction in these devices further limit their degree of programmability.Furthermore, such devices are not future-proof, as the device configuration does not support integrating multiple inputs when expanding to heterogeneously connected systems.In view of these challenges, there is a need to develop multi-terminal devices (with versatile gating options) capable of the synaptic programmability previously shown in nanoscale memristors.
Multi-terminal transistors with novel semiconductors can be gated using conventional dielectrics, ionic capacitors and optically using light pulses.Interestingly, dynamics in various modes can easily be coupled together with the semiconductor channel as the platform.The use of ionic capacitors with complex resistivecapacitive coupling that is highly dependent on device geometry produces complex dynamical phenomena that can be utilized in spatiotemporal processing of signals.This is especially important when multiple in-plane gate electrodes connected to the same semiconductor channel can perform not only spatial recognition but also encode temporal information into readable electrical signals.Furthermore, when coupling the slower dynamics of ionic gating with fast dynamics in electrostatic gating, complex dynamical effects such as homeostasis in biological systems can be emulated.When coupling optical stimuli with ionic-bonded semicon-ductor systems, ion-mediated recombination dynamics can also be utilized to produce synaptic plasticity with temporal characteristics distinct from regular phototransistors.Recent advances in semiconductors also uncovered materials that are responsive to stimuli and exhibit adaptive behavior through phase or redox changes.Redox-based semiconductors also show much higher tunability in electrical and optical properties as compared to conventional semiconductors.
In this review, we first categorize two distinctly pivotal approaches (Figure 1) to utilizing multi-terminal devices as (I) Coupling sensory inputs by dynamical material properties (Section 2) and for the (II) Efficient data sensing and signal processing (Section 3).Reconfigurable devices that employ coupled configurations that tap on multiple stimuli modes, electrode arrangements and logical circuit designs are explored.Subsequently, the efficient processing of sensory signals enabled by adaptive materials and sensorimotor circuits that assure quick response and reaction are explored.These multi-terminal transistor devices that closely emulate bio-realistic higher-order functions with low circuit complexity are expected to be vital components in nextgeneration bioinspired electronic and robotic applications, including humanoids, social robotics, and cybernetics.

Coupling Sensory Inputs by Dynamical Material Properties
Learning in the brain is a product of temporary and permanent changes to the strength of synaptic connections between the neurons.Analogous to the strength of connection in these synaptic junctions, the device conductance in transistor devices can be deemed as an adjustable synaptic weight.In Figure 2a, a typical transfer characteristic (I DS -V GS ) of a transistor is shown where the applied gate-source voltage (V GS ) modulates the conductivity of the transistor channel.Interestingly, transistors with memory can be designed as we will explore in this section of the review article.In these memory transistors, the channel conductance can be modulated either with temporary or permanent changes with timescales from milliseconds to minutes.Due to the slow discharging of ion-based capacitors, electrolyte-gated transistors tend to have volatile memory effects induced by remanent charges.However, due to the possibility of ion intercalation or creation of oxygen vacancies induced by the strong electric field, non-volatile memory effects can be programmed.This is reflected in the shift of transfer characteristics in the transistor (see Figure 2a).In studies that emulate synaptic plasticity, [21][22][23][24] the gate and source-drain terminals are analogous to the pre and post-synaptic neurons respectively (see Figure 2b).The synaptic weight is represented by the channel conductance (I DS read @V GS = 0 V) and programmed by gate voltage pulses (V GS ) modulating the transfer characteristics in the transistor device.The memory characteristics achievable would depend on the type of transistors used; this is determined by the semiconductor material and the capacitor used.In Figure 2c, a generalized categories of transistors are illustrated.The conventional thin-film transistors initially comprise of a semiconductor channel (metal oxides, small molecules, and conjugated polymers were used) and is gated by a dielectric capacitor (SiO 2 , high-k dielectrics).In order to induce a stronger capacitance, a new type of capacitor  using electrolytes (ionic liquids and polymeric gels) was introduced.The formation of electric double layers under an applied field induces extremely high capacitance at the semiconductorelectrolyte interface, accumulating high concentrations of electrons in the semiconductor channel.Ionotronic transistors using ionic conductive electrolytes including ionic liquid, ionic gel, polymer electrolytes, etc., have demonstrated good electrostatic modulation characteristics and unique interfacial electrochemical process, with great potential for incorporation in energyefficient neuromorphic devices. [25]The huge electrical doublelayer (EDL) capacitor (>1 μFcm −2 in thin film with a thickness in the order of 1 nm) in the ionotronic electrolyte induces high carrier concentration (>1013 cm −2 ) in the channel at low working voltages (<3 V). [26] In general, change in channel conductance which gives rise to memory effects, arises from electrostatic modulation and electrochemical reaction either at the interfaces or in the active material. [27]Using an example of an accumulationmode n-type semiconductor (e.g.TiO 2 ), when positive spikes are applied on the gate electrode, cations will migrate towards the electrolyte/channel interface, forming an electric double layer.The effective double layer accumulates electrons in the channel (akin to "doping") and induces a conductance change (electrostatic modulation).The memory retention of the system depends on the capacitive discharging rate of the double layer In the case of an electrochemical transistor or organic electrochemical [28] transistor, the redox activity between the electrolyte and semiconductor was thought to result in the doping/de-doping of the channel; the added chemical capacitance is capable of inducing longer-lasting memory retention than the former electric double layer transistor.There is an ongoing discussion about whether some of these architectures can be termed electrochemical transistors.
Tunable Volatile Memory with Single Gating: While the review article focuses on multi-terminal transistors, it is noteworthy that kinetic time constants of volatile memory achieved by single gating can be varied by changing the gate voltage and stimulation frequency; the time constants depend on the conductivity state of the semiconductor as well.The tunability can be attributed to the non-equilibrium charge imbalance in the electrolyte-semiconductor capacitor. [29]Furthermore, when using a mixed-ionic-electronic conductors like PEDOT:PSS, the use of a DC bias offset at the gate electrode (which governs transconductance) can finetune transient memory effects from depressing to potentiating behavior.This can be attributed to a possible electrostatically stored charge and a pseudo-inductive contribution by the electrolyte-PEDOT:PSS interface. [30]on-Volatile Weight Updates in Transistors: Under a static, equilibrium condition, the conductance of a semiconductor can reliably be determined based on the materials properties -semiconductor mobility, device capacitance, and effective gate voltage.In order to achieve long-term memory, the active material has to be capable of storing charge or undergo a chemical transformation.In the case of a transition metal oxide material, the MoO 3 /Li x MoO 3 is a great example of a stable chemical transformation. [31]The more conductive reduced form (Li x MoO 3 ) is stable and capable of retaining its state for a long time.The concept is similarly employed in energy-efficient electrochromic windows.Likewise, a chemically evolvable polymer that can undergo in-situ electropolymerization [32,33] has been shown to greatly improve state retention by two orders of magnitude.
While basic synaptic plasticity simply refers to modification in the strength of connectivity between two neurons, the realization of an interconnected neural network requires emulating the coupling between multiple synaptic devices.Their modulation should not be individual but be governed by the overall activity in the system, i.e., hetero-synaptic plasticity (the codependence between multiple synapses in the same junction) and homeostasis (the regulatory mechanism in the body meant that each local synaptic activity is affected by global oscillatory factors such as ion concentrations and other neuromodulators).The emulation of these mechanisms is critical when integrating multiple synaptic devices in a larger interconnected network.A multi-terminal device configuration would be better suited to handle the various input possibilities for modulating synaptic conductance.Dendritic integration through multi-in-plane gating utilizes spatially anisotropic capacitive coupling across the device geometry to modulate conductance, while synergistic multimodal gating induces temporally-distinct modulations of carrier concentration by tapping onto various activation modes -electronic, ionic, and photonic.In both these approaches, the final response becomes a convoluted function of the number of activated gates/modes of operation, thus increasing the weighted plasticity, memory storage capacity, and enabling the emulation of hetero-synaptic plasticity and homeostasis at the single device level.

Multiple Modal Inputs -Synergistic Electronic, Ionic, and Photonic Gating
Combining multiple gating controls that can induce temporallydistinct modulations of the channel carrier concentration would increase the weight plasticity and memory storage capacity, enabling emulation of more complex neuronal behavior.Following this approach, several reports employing electronic, ionic, and photonic controls have been recently reported with a focus on the degree of achievable plasticity.By perturbing the same active element via independent input schemes, these configurations allow for the activation of the ionic, electronic, or photonic mode, enabling higher order of plasticity to be achieved on the same device.
Comprehensive emulation of neuronal dynamics (e.g., homeostasis, dendritic integration, synaptic competition) demands effective modulation of the local post-synaptic strength in response to local pre-synaptic action potentials.However global modulatory factors also need to be considered.Large groups of biological neurons receive common inputs in the form of neuromodulators such as dopamine, noradrenaline, or acetylcholine, which correspond to reward or surprise signals in biological networks. [34]hus modulating learning rules through such global factors need to be considered-however, most studies only focus on the emulation of plasticity due to local temporal correlations.Global neuromodulations such as homeostatic mechanisms are often overlooked in artificial neural networks, limiting the attainable plasticity.By combining multiple gating controls to enable temporally-distinct carrier concentration modulation, both local and global oscillations can be addressed.Aside from the .Reproduced with permission. [35]Copyright 2017, Springer Nature.possibility of emulating more complicated neuronal behavior with fewer devices, power consumption, circuit area footprint and circuit density requirements of artificial neural networks can be alleviated.

Ionic-Electronic Gating
Global Versus Local Modulation: Gkoupidenis et al. demonstrated a seminal work on an OECT array with a global electrolyte and an array of immersed two-terminal devices, in which each device consists of a PEDOT: PSS channel (Figure 3a).By applying a voltage to the NaCl or DI water-electrolyte, the channel weights can be globally modulated, enabling the emulation of homeoplasticity (Figure 3b-d).The synchronization of I/O transmission and a global clock behavior achieved by leveraging this effect is then obtained by using local pulsed/modulatory and global (pulsed) oscillatory inputs, respectively.Furthermore, the strong capacitive coupling of the electrolyte enables the establishment of soft lateral connections between the individual devices, facilitating facile emulation of spatiotemporal homeostasis and coincidence detection.Homeostatic plasticity is a higher-order phenomenon, enabled by the convoluted charge imbalances induced by two different modes.The complex interplay between dual gating (ionic and electronic) was previously explored in pentacene ultrathin film transistors.The dual-gated architecture allows the authors to estimate the double layer capacitance in the 0.1 m NaCl aqueous medium to be in the range of 10 μF cm −2 . [36]velopment of such soft globally connected architectures via electrolyte gating paves way for the realization of neural networks of higher complexity with minimal hardwire connectivity. [35]nterestingly, this can also be realized by CMOS compatible processes.A dual-electronic-gated artificial synapse based on lowtemperature polysilicon (LTPS) was recently proposed as a scalable CMOS-compatible biomimetic synapse.The p-type LTPS thin film, dual-gated transistor was fabricated and addressed using an active matrix, a mainstream technology for flat-panel and flexible displays.SiO 2 and SiO 2 /Si 3 N 4 layers serve as the global bottom and local top gate dielectrics, enabling emulation of both excitatory and inhibitory modes via simple charge trapping mechanisms.Application of positive voltage on the top gate bends the energy levels of polysilicon down toward SiO 2 , trapping electrons at the SiO 2 /polysilicon interface.The depletion of electrons in the channel increases the conductivity of the p-type TFT, resulting in excitatory behavior.On the contrary, the negative voltage on the top gate traps holes at the interface, resulting in current inhibition in the p-type channel.With the help of a global modulatory bottom gate, the strength of excitatory and inhibitory currents was dynamically modulated, enabling the realization of homeostasis within the system. [37]omeostatic Plasticity: John et al. demonstrated an electronicionic dual gating approach where a solid-state ionic electrolyte and silicon dioxide (SiO 2 ), function as the local top and global bottom gate dielectrics, respectively.By capacitively coupling indium-tungsten oxide (IWO) semiconductor interfaces, this dual-gated architecture (see Figure 4a) successfully emulates Emulation of homeostatic and hetero-synaptic plasticity via ionic-electronic gate coupling.a) Schematic of dual-gated inorganic synaptic transistor. [38]b) Controlled facilitation and depression with dual gating approach.c) The dual-gating approach augment the plasticity and enable higherorder temporal correlations.Reproduced with permission. [38]Copyright 2018, American Chemical Society.
complex neuronal functions, such as hetero-synaptic plasticity, homeostasis, association, correlation, and coincidence.Ionic migration-relaxation kinetics in the ion gel (ionotronic mode), which mimics the influx of Ca 2+ ions in the dendritic spines, modulates the electronic conductance-state of the semiconducting channel.This process is in addition to the carrier trappingdetrapping phenomenon occurring at the IWO-SiO 2 interface (electronic mode) where the charge transport pathways emulate the synaptic cleft, and channel-conductance represents the synaptic weight.The ionotronic mode (top ionic gate) captures the effect of local activity correlations, while the electronic mode (bottom electronic gate) represented by global regulations, facilitates homeostatic regulation at an elemental level.With the coexistence of electron trapping and ion migration effects, the dynamic range of synaptic conductance is extended to 4.5 bits per synapse, thereby increasing the number of accessible memory states.Homeostasis (Figure 4b) and hetero-synaptic plasticity (Figure 4c) are implemented via additive/subtractive operation of the local and global gates.By perturbing a very thin semiconducting channel via two separate gates for the execution of Pavlov's dog experiments, fundamental rules such as associative learning and classical conditioning are demonstrated. [38]

Electronic-Photonic Gating
Photonic synapses utilizing optically-programmable active elements have recently attracted considerable attention due to the range of advantages offered -high bandwidth, low crosstalk, and ultrafast propagation speed. [39,40]In photonic synapses, the synaptic weight is altered by optical inputs instead of electrical signals.Compared to electrical writing, photo-programming provides an alternative for accessing such multi-level memories. [41]onfiguring memristive devices with optically and electronically modulatable conductance had promising results in high update linearity and low write noise, both of which are analog device properties that determine the accuracy of neural networks.This could be attributed to the precise delivery of power with light pulses as compared to stochastic switching in filamentary memristors.This entails the need for atomically-thin semi-conducting channels with opto-electronically modulatable carrier concentration, and synergistic gating strategies to address the memconductance states.Recently, studies investigating lightmodulatable synaptic plasticity features have been reported based on amorphous oxide semiconductors (AOSs), [42,43] 2D transition metal dichalcogenides (TMDCs), [44,45] quantum dots, [46,47] halide perovskites, [46][47][48] and organic or polymeric materials. [49]The persistent photoconductivity (PPC) phenomenon, which banks on trap-assisted delayed recombination of electrons and holes, is often utilized as a proxy to depict analog conductance transitions.While many studies directly perturb the conductance of the active layer using photo-active switching semiconductors, others utilize photo-absorptive capping layers to indirectly modulate memresistance of the active layer.Such demonstrations open up the possibility of using a photo-electrical hybrid approach to enhance the computational functionalities.
Photonic-Only Potentiation and Depression: Most photonic synapses studied to date require an electrical bias to erase the states or in other words, induce depression.However, Qin et al. demonstrate a unique photonic depression behavior in a hybrid phototransistor achieved with the aid of a gate bias.With the design based on a graphene and single-walled carbon nanotube (SWNT) heterostructure, this device exhibits switchable doping (p/n) behavior as a function of the gate voltage, enabling the observable transition between the potentiation and depression modes.Optical illumination below a particular threshold voltage (V cross ) leads to electron injection into the p-type graphene, which reduces conductivity and promotes depression.On overcoming this threshold voltage (V cross ), the device exhibits n-type semiconductor characteristics, where on injection of additional electrons via optical illumination, an increase in conductance or potentiation, is observed (Figure 5a,b). [50]xide Semiconductor Based Photo-Active Devices: As a representative example, Lee et al. demonstrated photonic neuromorphic devices using AOSs with a high degree of plasticity.All major synaptic functions, including STP, LTP, and STDP are realized as a function of the wavelength, intensity, frequency, and the number of optical stimuli across a variety of material platforms including indium-gallium-zinc-oxide (IGZO), indiumstrontium-zinc-oxide (ISZO), indium-strontium-oxide (ISO) and Reproduced with permission. [50]Copyright 2017, IOP Publishing.
indium-zinc-oxide (IZO).These synaptic functions are successfully emulated by harnessing the inherent PPC phenomenon, generally believed to stem from the ionization of oxygen vacancies.In this study, the activation energy for neutralization of ionized oxygen vacancies is noted to play an important role in determining the excitation and relaxation dynamics. [51]ang et al. demonstrated photoelectrical devices in a fieldeffect configuration with In-Zn-O channel and ion gel dielectric (Figure 6a).The inherent PPC of the In-Zn-O films permits dynamic regulation of synaptic plasticity through exposure to deepultraviolet (DUV) light.The significant increase in photocurrent with increasing light intensity from 18.23 to 364.53 μW cm −2 indicates its high sensitivity to DUV light.A wavelengthdependent study reveals that the strong excitatory behaviour observed at lower wavelengths arises from the generation of larger number of electron-hole pairs.In addition, the pre-synaptic light  [42] d) A 3D schematic illustration of CsPbBr 3 QD-based flash memory.e) Example of photonic potentiation and electrical habituation implementation.Reproduced with permission. [42]Copyright 2018, IOP Publishing.Reproduced with permission. [52]Copyright 2018, John Wiley and Sons.
spike (1.82 mW cm −2 , 100 ms) stimulated weight changes (Δw) exhibit both gate (V GS ) and drain voltage (V DS ) dependency, enabling metaplasticity to be achieved in the device (Figure 6b,c).Weights can be tuned by modulating the Schottky barrier across the In-Zn-O/ion-gel interface through changes in amplitude of V DS , V GS , and the light inputs.The polarity of the gate voltage activates electron/hole trapping centres at the interface in turn modulating the recombination rates of photo-activated electronhole pairs, manifesting in the form of weight changes.The implementation of STDP learning principles on this photoelectrical device via photonic potentiation and electrical depression fulfils the switching requirement needed for efficient computing. [42]Similar gate-bias dependent behaviors observed in indium-galliumzinc oxide (IGZO)-based thin-film transistors (TFTs) have also been reported by Wu et al. [43] Transition Metal Dichalcogenide Photo-Active Devices: 2D TMDCs with their atomically-layered structure [53] and intriguing optical-electric properties [53,54] present a platform to circumvent the limitations of traditional semiconductors in accordance to Moore's law.Recently, neuromorphic transistors based on 2D TMDCs exhibiting interesting opto-electronic properties are realized utilizing the photoelectrical hybrid modulation approach. [44]e et al. demonstrated an ultrathin memristive synapse based on monolayer n-MoS 2 /p-Si heterostructure.Versatile synaptic neuromorphic functions, including STM, LTM, PPF are successfully mimicked based on the inherent PPC of MoS 2 and its volatile resistive switching behavior.The PPC effect in MoS 2 is confirmed to originate from random localized potential fluctuations and resistive switching behavior is believed to stem from the trapping of electrons at the MoS 2 /SiO 2 interface.These effects are harnessed to create photonic potentiation and electric habituation schemes.The increase in conductivity is attributed to the generation of electron-hole pairs under light illumination, while depression is achieved by the application of negative bias on the back gate, which consequently traps electrons at the MoS 2 /SiO 2 interface and decreases conductivity. [44]Jiang et al proposed a similar photoelectrical hybrid memristive device comprising of 2D MoS 2 channels and a sodium alginate biopolymer dielectric.The potentiation filtering effect is realized with the help of an electrical gating approach, while pulse frequency photonic modulation enables the implementation of both potentiation and depression filtering effects.This photonic depression filtering effect is related to the trapping of photoelectrons at the MoS 2 /sodium alginate interface.Spatiotemporal correlation effects and STDP are both achieved by utilizing the synergistic interplay between electrical and photonic stimuli. [45]uperlinear Weight Updates: Some key considerations for deep recurrent neural networks include high precision weight update, high update linearity, and low write-noise, to impart the capability of addressing complex tasks such as speed recognition and natural language processing, which require learning of temporal signals.The degree of weight precision required in deep recurrent neural networks is challenging for conventional memristive devices due to their abrupt switching dynamics.Drawing inspiration from optogenetics -a neuromodulation technique using optical pulses to modify neuronal activities in the brain, photo-stimulation may be utilized to selectively modulate the synaptic weights of neuronal arrays.Recently, by synergistically combining both electrical and photonic modulation,  [67] b) Highly modulatable weight changes with a combination of two or more modes -electrostatic, ionotronic, and photoactive.Reproduced with permission. [67]Copyright 2018, John Wiley and Sons.
enhanced under optical excitation.The conductance transition and on-off ratio increases under green light and UV illumination due to the higher trapping frequency of photo-generated holes in the BP-ZnO NP.A large negative pulse is applied at the gate as a pre-programmed pulse to ensure sufficient trapping of holes in the BP-ZnO NPs.Metaplasticity effects in the device are obtained by leveraging the complementary operations of the electronic and photoactive modes. [63]hotosynthetic Proteins Sensitized Photo-Active Devices: Photosensitizers have been demonstrated as a versatile method to functionalize semiconductors with photo-activity.In addition to the well-studied photovoltaic absorber thin films, photosynthetic proteins [64][65][66] with their high quantum yield and the ability to store energy can provide great memory device properties.In these works, Rhodobacter sphaeroides are a type of bacterium with a reaction center consisting of three-subunit transmembrane protein.The multi-protein complex is capable of converting light energy into chemical energy, via the photosynthetic process.Instead of converting light information into electrical signals (as regular electronic transducers do), the absorption of light generates charged pairs (a primary donor P + and a final acceptor Q B − ) which are separated and stored as chemical potential energy.Photosynthetic apparatus constructed using films of proteins have been shown to generate sub-1 V of photovoltage and capacitance of >10 μF cm −2 . [66]The prolonged charge storage capabilities, when coupled with a semiconductor such as PEDOT:PSS, can provide unique information storage capabilities. [64]

Coupled Electronic-Ionic-Photonic Gates
John et al synergistically combined the aforementioned gating controls, namely -electronic, ionic and photonic, to realize neuromorphic elements demonstrating metaplasticity, or plasticity of plasticity (see Figure 7a). [67]Utilizing trapping-detrapping (electronic mode), ion migration-relaxation (ionotronic mode) and persistent photoconductivity (photonic mode) mechanisms at the semiconductor-dielectric interface, artificial synapses are realized in a MoS 2 field-effect transistor configuration.Due to these different mechanisms, diverse spatiotemporal properties coexist on the same device.Symmetric anti-Hebbian STDP protocol is realized with the help of electronic-mode operation, while an ionotronic-mode operation results in a symmetric Hebbian STDP protocol.Moreover, the individual gating controls can be combined to realize metaplasticity and homeostasis on the same device, a feat impossible/challenging with conventional 2terminal memristors.Additive/subtractive operation of the various modes offer precise control over both positive and negative weight changes (facilitation and depression, respectively) (see Figure 7b), metaplasticity and homeostatic regulation.Implementing classical conditioning using unconditioned optical and conditioned voltage stimulations, results in the effective association between the operation modes.In conclusion, these comprehensive results portray the advantages of a multi-modal-input architecture and paves the way for biologically-plausible metaplastic architectures for advanced neuromorphic computing. [67]igure 8. Dendritic integration and spatiotemporal processing with multi-electrodes system.a) Schematic representation of neuron with starch-based dendritic transistor, b) integration of two dendrites and the spikes corresponding to EPSC triggered by two input spikes V 1 and V 2 with 1.5s interval with non-linear integration.Reproduced with permission. [74]Copyright 2018, American Chemical Society.

Multiple In-Plane Gate Electrodes
In contrast to two-terminal memristors, external gate modulation capability in three-terminal memristors enables the decoupling of read (reading the channel conductance via the source-drain electrodes) and write (channel conductance modulation using gate terminal) operations, thereby allowing internal signal transmission and modification to be performed simultaneously without the need for external circuitry. [68]In these cases, the capacitive coupling becomes critical with various mechanisms such as electrical double layer (EDL) effects and ferroelectric switching having been explored and successfully exploited to effectively gate the channel memconductance. [25][69][70][71][72][73] In this section, we will focus on various studies utilizing multiple in-plane gates for realization of spatiotemporal dendritic integration.The integration of excitatory and inhibitory dendritic inputs onto memristors or "dendritic integration", is a key feature of hetero-synaptic plasticity -the horizontal communication at a junction between multiple synapses.Dendritic integration between the transistor devices can be achieved by adopting a multiin-plane gating configuration with anisotropic spatial capacitive coupling.
As compared to two-terminal memristors, synaptic transistors are biologically similar to dendritic synapses as their multiterminal nature allows for simultaneous processing of spatiotemporal input spikes and offers the capability for dendritic integration [69] and hetero-synaptic plasticity.Simultaneous operation of multiple in-plane gates allows precise and superior con-trol over channel conductance modulation and advanced synaptic functions like logic operation, dendritic integration, sub, and super-linear summation, [25,[67][68][69] respectively.In addition to capacitive coupling capable of emulating basic interconnectivity between multiple synaptic elements, a more bio-realistic representation of hetero-synaptic plasticity with self-adaptive learning ability can also be achieved by limiting ionic doping in the system.Since the ion dynamic characteristics (concentration-driven and diffusivity) are similar to competition for limited resources in synapses, the activity of one synapse can greatly influence those of neighboring synapses in an ion-mediated switching material system.This enables the implementation of hetero-synaptic plasticity to construct a self-adaptive neural network with high synaptic stability.In this section, the dendritic integration and heterosynaptic plasticity implemented in multiple in-plane configuration as well as progress in the field of ionotronic devices towards the adoption of self-adaptive synapses will be discussed.

Dendritic Integration
Gao et al demonstrated a synaptic transistor with two dendritic gates, designed to achieve basic logic operations like OR, AND, and AND-NOT. [69]In artificial synaptic networks with dendritic integration, the interconnected gate electrode or channel acts as the dendrite, whereby the application of two pre-synaptic spikes can trigger post-synaptic current through an interconnected gate as shown in Figure 8a.Asynchronous firing or absence of any of the two dendritic inputs leads to an "OR" logic operation, with EPSCs failing to overcome the activation threshold.However, with the synchronous firing of the two dendritic inputs V 1 and V 2 ,  [70] the cumulative EPSC increases to a level above the threshold, enabling "AND" logic operation to be demonstrated (Figure 8b).Interestingly, the arrival of synchronous inputs at the two dendritic gates results in a superlinear output response > sum of individual outputs analogous to CA1 pyramidal neurons. [69]A similar behavior was also observed by John et al using ZnO semiconducting channels gated by a [EMIM][TFSI]-PVP ion gel. [75]Such superlinear behavior can be attributed to ion migration-relaxation dynamics in ionic dielectrics with systematic, in-depth investigation required to exploit these unique features for real-life computing applications.
Recently, Fu et al. built a 9-element (3×3) interconnected, flexible 3D artificial neural network based on multi-terminal poly(3hexylthiophene) (P3HT) organic electrolyte-gated transistors. [67]everal pairs of source and drain electrodes were deposited on the P3HT channel; the lateral coupling of the device multiple gate electrodes is achieved using an ion-gel composed of [EMI] [TFSA] and PVDF-co-HFP (Figure 9a).This multi-gated configuration boosts the overall weight change and allows precise modulation of long-term memory via synergistic effects, enabling basic functions such as potentiation/depression and STDP to be implemented with great flexibility and reliability. [67]Moreover, the configuration offers responsivity to spatial differences in input spikes, where changes in input orientation is translated to differences in gate-to-channel distance and manifests in the form of excitatory post-synaptic current (EPSC) amplitude change.A similar approach was taken by Qian et al, where a multi-gated array, consisting of poly (3-hexylthiophene) P3HT organic electrolyte-gated transistors and ionic-gel (P(VDF-HFP) + [EMI][TFSA]), was used to demonstrate spatiotemporal dendritic integration.The EPSC can be modulated by varying the dendrite-post synapse distance and number of pulses applied.The strong dependence of capacitance over the individual pixel lateral distances resulted in differentiated transistor characteris-tics in terms of both their temporal responses and overall signal gain.The lateral capacitive coupling, utilized as a proxy to demonstrate orientation-dependence is shown in Figure 9b.As observed, for large EPSC response to be induced, a longer pulse duration and short dendrite-post synapse distances are required.Similarly, the PPF exhibits a strong dependence on spatial capacitive coupling variations and is demonstrated as a function of dendrite-post-synaptic distances and spiking interval. [70]To realize orientation-based tuning in artificial synapses, akin to that present in the primary visual cortex, two spatially-oriented lateral inputs are assigned to gate 0 and gate x (numbered 1-6), resulting in the modulation of EPSCs as a function of the orientation (0°t o 180°).This multi-gated architecture also allows one gate to be utilized as a modulatory terminal while the other gates function as input pre-synaptic terminals, enabling additional logic implementations and weight modulations to be executed on the same device. [70]

Hetero-Synaptic Competition
While capacitive coupling is capable of emulating basic interconnectivity across the multiple synaptic elements, a more biorealistic approach for dendritic integration involves imparting ionic competition behavior in the hetero-synaptic junction.Ionic dynamics, characterized by its diffusivity and concentrationdriven nature, can be leveraged to fully emulate the dynamics in biological synapses.Unlike earlier reports where ionic dynamics were emulated using software or voltage-divider circuits, Zhu et al. have shown success in utilizing the highly anisotropic and low activation energy ion transport properties in 2D transition metal dichalcogenides. [77]In this work, the competition for Li + ions in the hetero-synaptic device is analogous to that of plasticity-related protein (PRP) in a biological neural network The competition for Li + in the channel is analogous to the limited proteins in biological neural networks.Reproduced with permission. [77]Copyright 2018, Springer Nature.
(Figure 10a).The aforementioned plasticity-related proteins are limited synaptic resources necessary for long-term weight changes, meaning depression activity in one synapse will facilitate the potentiation activity of another.To demonstrate synaptic competition, a coplanar multiple in-plane terminal device with Li + -intercalated MoS 2 channel was fabricated to emulate a heterosynaptic junction.The high in-plane Li + diffusivity in MoS 2 provides ionic paths with low activation energy suitable for transmission of Li + ions across the distance between the individual synapses.Furthermore, the phase transition between semiconducting 2H polymorphs and metallic 1T' can be induced by the intercalation of Li + ions, ensuring reliable electrical switching.To further understand the role of Li + ions in resistive switching, X-ray photoelectron spectroscopy was used to verify the change in the chemical state of MoS 2 after lithiation.In the lithiation process, Mo 4+ was reduced to Mo 3+ to accommodate the Li + into the lattice.Raman spectroscopy illustrates the reversible phase transition from semiconducting 2H to metallic 1T' when the device switches from high resistance state (HRS) to low resistance state (LRS) and back.The expected volume expansion after lithiation arises from the increase in lattice spacing from 0.62 to 0.71 nm observed via cross-section High-Resolution Transmission Electron Microscopy (HRTEM) imaging.The comprehensive chemical characterization verifies the role of field-controlled Li + migration during reversible electrical switching, without dismissing the contribution of localized charge trapping to the memristive behavior.Modeling of electrical IV data further indicates that resistive switching is facilitated by the conversion of Schottky contacts to Ohmic contacts at the Au/Li x MoS 2 interfaces.In the 4 pre-synaptic terminal configuration (see Figure 10b), the presynaptic terminals can be stimulated individually, while the post-synaptic terminal is assigned to the ground.In this manner, all 4 synapses can share and connect with the same Li x MoS 2 channel.For the initialization step, one of the electrodes can be stimulated with negative voltage to induce loss of Li + ions in the corresponding synapse.This process leads to the diffusion and accumulation of Li + ions in the neighboring synapses.Since the transition of HRS to LRS state involves the intercalation of Li + ions, the neighboring synapses will undergo a higher conductance change when stimulated with a positive voltage subsequently.Such interdependence of synaptic activity is mutual, providing a demonstration of ionic synergy when the activity of one synapse facilitates or suppresses the activity of another.Such synaptic competition or cooperation behavior is important for the stability of neural networks, similar to homeostasis-regulated plasticity.The actual demonstration of ionic dynamics (concentration-limited and diffusivity) to emulate the competition for plasticity-related proteins will pave the way for the implementation of more bio-realistic hetero-synaptic dynamics in a coplanar configuration.
Large-Scale Fabrication and Integration: Given the success in multi-terminal memristors and memtransistors, Won et al. demonstrated large-scale integration of multiple neuronal connections. [78]The experimental implementation of proposed Spiking Neural Network consists of 3 neurons with 9 synapses each.The physical hardware could classify between 3 classes -"vertical line", "horizontal line" and "orthogonal line".While elementary, the work demonstrates fabrication capabilities of monolayer graphene.Similar fabrication capabilities with transition metal dichalcogenides [79] were also demonstrated by other research groups, showcasing the feasibility of such technologies.In addition to challenges in transferring these layered materials, the stability to oxidation and humidity of the emerging dichalcogenide materials is also a concern.At this front, oxide encapsulation layers deposited by atomic layer deposition [80] or a metallic chromium layer have been shown to improve the stability of these materials.

Efficient Data Sensing and Signal Processing
While memristive devices have been utilized in several highperformance computing applications such as in-memory computing and hardware accelerators for deep learning, it is challenging to meet the requirements of high weight precision and conductance linearity.Over the last few years, extensive progress in the development of memristive devices has opened new avenues that require only a low-to-medium degree of computational precision.The integration of sensors with memristive elements promotes the development of biomimetic perception systems.With the increasing complexity of sensory elements in a system, the amount of data transmission becomes increasingly inefficient as well.There is a need to reduce the amount of redundant data, by imbuing memory elements closer to the sensory elements.This compelling area of research has received significant attention from the scientific community due to the potential for development of intelligent perception systems with inbuilt learning ability.
Biological computing generally learns and adapts from inputs to the central nervous system via an array of pressure (skin), visual (eyes), sound (ears), olfactory (nose), and taste (tongue) sensors.Creating artificially intelligent, interactive robotic systems calls for a similar approach and entails the integration of multiple sensory elements.This can generally be achieved either with stimuli-responsive materials that can respond adaptively or incorporating memory elements with the sensory devices.Devices that depict modulation of memresistance states via multiple gating controls open up new possibilities for integration with sensor stimuli such as pressure, light, sound etc., further motivating investigations in this area.Such sensory information can be directly introduced to the neuromorphic circuits in the form of gate input, enabling learning of sensory data and enhancing sensor properties such as responsivity and selectivity to specific stimuli.Therefore, the platform can be trained simultaneously, with lasting modification to the sensory behavior required of each environmental condition.As such, the performance in achieving perceptual tasks is enhanced by preceding experience, and eventually exceeds the functionality of conventional sensors.
Demonstration [67] of such systems not only sheds light on sensory information processing but also enhances the performance of complex sensing applications such as tactile perception in electronic skin and machine vision.Electronic skins can be equipped with a self-learning perception of temperature, pressure, and other sources of stimulation and exhibit actions and reactions such as reflexes, yielding more accurate actuation, particularly for robotics.Recently, researchers have attempted to mimic the complex behavior of vertebrates via artificial circuits for nextgeneration bio-inspired robotics and humanoids.Unlike conventional sensors, which are passive and linear, biological perception systems can adapt to input changes and optimize accordingly depending on the environment.In this section, recent studies on stimuli-responsive materials, sensory gated (light/tactile) synap-tic transistors for realizing sensory-memory integrated circuits will be discussed.

Adapting by Adjusting to Environment
Many real-life sensory inputs are dynamic and abrupt in nature (e.g., a flood of light, sudden blackout, or extremely sharp objects), with a risk of not only damaging our sensitive receptors but endangering our lives.In the case where excessive activation would be undesirable (exceedingly high intensity), further stimulation must be suppressed or avoided.Likewise, when stimulation is weak, signal enhancement is required.In order to optimize sensory capabilities, the retina in the eyes together with tunable activation of cones and rods can adapt to various light intensities and switching of priority between differentiation and detail.In order to improve sensory performance, there is a need to implement the key feature of self-adaptation.Biological sensory systems rely on complex biochemistry to demonstrate signal transduction capable of adapting to the environment.In contrast, conventional sensor electronics are based on semiconductor materials that respond monotonously to stimuli.In order to emulate the biological systems, stimuli-responsive materials that can illustrate adaptation are key to improving the versatility of sensor electronics.
Sensory Adaptation with Electrochromic Stimuli-Responsive Sensing Materials: The visual sensory system is complex, it relies on biochemical reactions to sense the environment.In order to realize an artificial optic nerve connected to photoreceptors (rods and cones), the two functionalities (Signal transduction and Adaptation) must be emulated.Phototransduction in the eyes is the conversion of optical information into electrical readable signals at the photosensitive retina (see Figure 11a).Adaptation to light and darkness is enabled by the switching between the two major classes of photoreceptors (rods and cones).In low light conditions, rods are physiologically responsible for reducing the activation threshold to increase signal collection.As the time spent in the dark increases, the sensing threshold is actively lowered.Conversely, on exposure to high light intensity, the activation threshold is raised to improve visual contrast.However, the main challenge to developing an artificial retina is in imparting variable sensitivity (threshold) to these inputs.The passive nature of conventional photodetectors implies behavior or response is unaffected by light intensity.While photosensitive synaptic transistors exhibit some form of learning behavior (e.g., PPF, SRDP, associative learning), they are usually programmed with training pulses of similar light intensities.The activation threshold remains unchanged regardless of the dynamicity of the external inputs, leading to poor contrast particularly at extremely low or high intensities.In order to imbue adaptation functionalities, Ng et al. utilized an electrochromic oxide semiconductor (MoO 3 ) as the sensing layer gated by a Li + ionic gel. [82]The optical absorption properties of the sensing layer can be altered by the application of electrical pulses.In that way, the photoresponse of the transistor-based photodetector can be tuned in demand (see Figure 11b).The electrochemical modulation of optical absorption spectrum is similar to the photochemical switching between cone and rod photoreceptors in the eye.In the photopic mode with enhanced cone activity, the device exhibits high The degradation of sensitivity to red light and improvement with blue light when exposed to a train of pulsed stimuli.c) Sensory adaptation to dark as illustrated with the repeated cycles of stimuli and adaptation.Reproduced with permission. [82]Copyright 2022, Elsevier.
photoresponse across the visible spectrum (Red, green, and blue); this is critical for activities such as hunting and gathering during the day (see Figure 11c).In contrast, the scotopic mode with enhanced rod activity exhibits high selectivity toward blue light and higher signal-to-noise ratio; this is critical for danger detection during the night.While the electrochromic photodetector exhibits tunable photoresponse, the device is not self-adaptable and would manual electro-modulation.
Sensory Adaptation with Photochromic Stimuli-Responsive Sensing Materials: In order to illustrate the self-shiftable threshold, Hong et al. demonstrated a hybrid phototransistor, comprising a halide perovskite layer as the adaptive photosensitizer and MoS 2 channel as the charge transport. [81]The hybrid phototransistor device is exposed to a train of optical stimuli (1 mW cm −2 , 1 Hz) while the MoS 2 channel conductance is monitored continuously at V GS = −40 V and V DS = 1 V.Under repeated illumination, the phase segregation of the mixed halide perovskite CsPb(Br 0.5 I 0.5 ) 3 into the higher bandgap CsPbBr 3 results in the degradation of red light sensitivity while the sensitivity to blue light stimuli is enhanced.Interestingly, the phase segregation was shown to be reversible with relaxation time in the dark.This stimuli-responsive behavior of the mixed halide perovskite allows for the reconfigurability of the hybrid device similar to the biological sensory system.As a result, both the stimuli sensing and sensory adaptation to dark, depicted by the increase and decrease of photosensitivity respectively, can be emulated.In a second work by Kanwat et al., a photoconductive device with a <110> FA n+2 Pb n Br 3n+2 quasi-2D photochromic perovskite was employed as both the sensing and the adaptive layer (see Figure 12a). [83]Optical stimuli (10-120 mW cm −2 ) are introduced while the photocurrent is collected with a small reading voltage of 0.1 V.In order to illustrate the utility of stimuli-responsive materials (photochromic quasi-2D per-ovskite) in photodetection, Kanwat et al. demonstrated 2 photodetector arrays with 4 pixels each (see Figure 12b).The left array is covered with black cardboard and represents the dark-adapted eye of a pirate, whilst the right array is exposed to stimuli and represents the working eye.When a low-intensity light is introduced, the working eye exhibits scotopic vision and high clarity of the image is observed.When switched to a high-intensity light abruptly, the working eye undergoes flash blindness with poor vision.However, the photochromic perovskite device adapts to the highintensity light, shifting the threshold, recovering the clarity of vision in photopic vision.The transition back to low-intensity light would hence result in dark blindness; in this case, the schematic pirate would remove his eye patch and utilize his dark-adapted eye with excellent scotopic vision.Utilizing stimuli-responsive materials such as the electrochromic and photochromic semiconductors, the realization of structural reconfigurability in the sensing material enables adaptive sensing similar to the biological sensory system.
Sensory Adaptation with Integrated Device Circuit Configuration: In order to illustrate a sensor system with shiftable threshold, a simple solution involves a circuit consisting of a sensor and a load transistor (as a voltage balance) coupled with a synaptic transistor.Kwon et al. demonstrate an artificial visual perception system [84] capable of adapting to different light intensities (see Figure 13a).Here, the CdSe photosensor is responsible for detecting the light intensity and converting it into a spiking voltage via a voltage divider.A higher light intensity will result in a higher spiking voltage to stimulate the IGZO synaptic transistor.Most importantly, the threshold can be tuned by modulating the load voltage, which is directly responsible for tuning the effective spiking voltage (see Figure 13b).When the response is high (photopic mode), due to high light intensity, the load voltage can be increased to ≈0.24 V Figure 12.Sensory adaptation with photochromic photodetectors.a) A schematic photoconductive device based on single-layer <110> quasi-2D perovskite acting as both the adaptive and sensing layer.b) By emulating changing lighting conditions with a series of low and high intensity light stimuli, the respective adaptation stages (scotopic vision, flash blindness, photopic vision, and dark blindness) are demonstrated.Reproduced with permission. [83]opyright 2022, American Chemical Society.
to improve the contrast.Under dim light conditions, when the response is low (scotopic mode), the load voltage can be lowered to 0.22 V to improve the contrast with a high decibel difference.The realization of a self-adaptive artificial optic nerve will pave the way to high-performance intelligent sensors capable of optimizing operating parameters according to the ambient environment.While self-adaptation is responsible for optimizing the collection of sensory information, perception is the organization and processing with the aid of learned information stored in memory.Perceptual learning is an important feature that can be achieved by integrating both the sensory and memory elements.

Sensorimotor Transmission between Sensor and Actuator
In addition to utilizing stimuli-responsive materials, sensory adaptation can be achieved with device configurational optimization.This involves the integration of the sensory element with memory devices; voltage-divider circuits are often used as a threshold shifting control in such implementations.The implementation of circuits like these allows for the utilization of highmobility and highly photosensitive semiconductors instead of limiting to a narrow class of stimuli-responsive materials.

Tactile Perception System
Recently, memristive devices have been integrated with various types of pressure sensors to mimic the artificial tactile perception.][87][88] Wan et al. [85] recently proposed an artificial sensory neuron (NeuTap) with tactile pattern recognition and perceptual learning abilities.This system involves the coupling of a highly-sensitive pyramid-shaped resistive pressure sensor (receptor) with a tungsten doped indium oxide transistor (synapse) via polyvinyl alcohol-based ionic cables (axon).The pressure sensor converts the pressure stimuli to electrical signals, which pass through a soft ionic cable to the neuromorphic transistor.This configuration resembles the working of an artificial biological sensory neuron as shown in Figure 14a.The neural response, realized by integrating two pressure sensors, is enhanced by a spatiotemporal correlation effect.The Neu-Tap is capable of recognizing tactile patterns, where flat and convex are deemed as ("0") and ("1") cases respectively (Figure 14b).The patterns are identified by the change in conductance as a function of pressure change, in the IWO synaptic transistor.All four possibilities of conductance are realized, namely, "11", "10", "01" and "00".The highest conductance change is observed for the "11" pattern due to application of consecutive pressure stimuli.A supervised machine learning method is implemented in the neuron to improve pattern recognition.The recognition error rate showed a dramatic reduction from 44% to 0.4% within just a few repeated training cycles (Figure 14c), akin to perceptual learning observed in humans.

Light Perception System
Similar to intelligent tactile-sensing devices, integration of photosensors with memristive devices have received considerable attention for applications such as image recognition and classification. [87,89]As a fundamental demonstration, Reproduced with permission. [84]Copyright 2019, John Wiley and Sons.c) Schematic of a sensory neuron compared to NeuTap.d) Tactile perceptual learning and pattern recognition by NeuTaP on finger.Schematic of a two-bit binary code for three pattern pairs along with the RI for four groups of training data.e) Error rate recognition as a function of learning times, where the measured data and line fitting are denoted by the black squares and red line, respectively. [85] [85]Copyright 2018, John Wiley and Sons.f) The modulation of the Venus Flytrap using spike-train from the artificial organic electrochemical neuron.Reproduced with permission. [90]Copyright 2022, Springer Nature.
Lee et al. [89] reported an organic optoelectronic sensorimotor artificial synapse using stretchable organic nanowire (thiophene diketopyrrolopyrrole (FT4-DPP)-based conjugated polymer and polyethylene oxide) synaptic transistors (s-ONWST) gated with an ionic gel (poly(styrene-b-methyl methacrylate-b-styrene) (PS-PMMA-PS) triblock copolymer and 1-ethyl-3-methylimidazolium bis(trifluoromethylsulfonyl)imide ([EMIM][TFSI])).A photodetector (heterojunction of ZnO and mixture of P3HT and PC 60 BM) converts the input optical signals to electrical voltages to drive the organic artificial synapses.The current output of the synapses is converted to voltage spikes by a transimpedance amplifier and input to a polymer actuator (PSS-b-PMB block copolymer) which acts as the artificial muscle fibre.To demonstrate the potential of these synapses as a route for optical wireless communication at the interface of human-machine interaction, the devices are subjected to simple images (light patterns) like "SOS" representing the International Morse code, and the EPSC responses were recorded as a function of the pattern of visible light.In addition to visible light, short messages ("HELLO" and "UNIVERSE") input through a silicon solar cell using infrared (940 nm) and ultraviolet (365 nm) light, provide a successful demonstration of recognition.Such sensorimotor synapses offer a promising strategy for the development of bioinspired soft electronics, neurologically inspired robotics, and electronic prostheses.

Spike Generation from Sensory Inputs
In order to fully emulate the sensory systems, spike generation from injected sensory currents is a key neuronal function.The neuron converts sensory information from the respective receptor cells into electrical spikes which can then be transmitted across the neural network (see Figure 15a).Currently, this function is usually emulated with digital hardware or power electronics.Recently, Harikesh et al. have demonstrated organic electrochemical transistors based artificial neuron [90] using an Axon-Hillock circuit (see Figure 15b).By injecting a constant current (a common protocol utilized by neuroscientists) into the neuron, a series of processes would result in a voltage spike (see Figure 15c).The voltage spikes will continue for as long as the current is introduced.Interestingly, the amount of injected current would determine both the amplitude and rate of the Reproduced with permission. [86]Copyright 2018, The American Association for the Advancement of Science c) Decentralized computational platform by embedding associative learning elements close to sensor nodes.d) A schematic on how the diffusive memristors are incorporated as nociceptive elements and drift memristors as synaptic elements in a decentralized computational platform. [92]ikes.An organic electrochemical synapse based on electropolymerization was also introduced in this work (see Figure 15d).With voltage spikes, synaptic modulation such as facilitation and depression can be emulated (see Figure 15e).By connecting the artificial neuron to an artificial synapse, they have also shown key characteristics such as spike-timing dependent plasticity (STDP).More interestingly, the team interfaced their artificial neuron to a Venus Flytrap plant (see Figure 15f).By injecting a current into the artificial neuron, the generated voltage spikes were introduced into the plant; as a result, they were able to "trick"/stimulate the plant to close.Very recently, Harikesh et al. have also realized a more biorealistic version with negative differential resistors based on the ambipolar ladder-type semiconductor poly(benzimidazobenzophenanthroline) (BBL). [91]Due to the ambipolar behavior, they were able to emulate the dual dynamics of fast sodium activation and slow sodium inactivation, therefore demonstrating key neuronal activation characteristics such as depolarization, repolarization, and hyperpolarization.
On encountering noxious inputs, our body can react via a reflex arc (e.g., partial closing of eyelids facing the sun, withdrawing of the arm under high pressure) to avoid further damage.While self-adaptation can optimize the sensory operation to suit the working environment, they are not connected to the motor response of the system.In face of noxious stimuli, the reflex arc reaction is critical in communicating rapidly to the mechanical actuation and steering the system away from danger. [92]

Reflex Arc -Sensorimotor Connection
A reflex arc is a form of untrained response that does not require any conscious effort.In electronic terms, instead of processing through a powerful central processor, it taps on a simple and lo-calized information processing loop to generate quick responses to situations such as escape and flight from danger.Kim et al. [86] demonstrate a flexible organic artificial afferent nerve as a type-I sensory neuron. [93]A hybrid bioelectronic reflex arc capable of controlling a biological motor nerve is demonstrated by connecting this organic artificial afferent nerve (AAN) to a cockroach's efferent nerve (see Figure 16a).The artificial organic afferent nerve consists of a highly sensitive carbon nanotube-based resistive pressure sensor connected to an artificial organic (pentacene) nerve fiber (pseudo-CMOS ring oscillator), which converts pressure stimuli into pre-synaptic voltage pulses for the organic neuromorphic transistor (Pentacene).Using this bioelectric reflex arc, changes in the frequency, intensity and stimulation of AAN produce prominent responses from the cockroach's muscle as shown in Figure 16b.In the event of exposure to a noxious stimulus, inference and decision-making process are accelerated (the latency between a stimulus and response is greatly reduced).Very recently, Wang et al. further advanced the monolithic integration of the low-voltage electron skin capable of neuromorphic sensorimotor loop. [94]Integration of such sensor and actuator elements onto neuromorphic circuitry would enable the development of intelligent robotics, sensory exoskeletons, and neuronal prosthetics.John et al. further expanded on the decentralized computational approach by incorporating nociception [95] (the detection of pain) into the reflex arc (see Figure 17a). [92]In this work, a diffusive transistor based on an ionic gel dielectric is deployed as an artificial nociceptor with gated-threshold switch function capability.Not only is the architecture simpler than a 6x transistor-1x capacitor configuration obtained via the Complementary-Metal-Oxide-Semiconductor (CMOS) fabrication technique, the ionic gel dielectric used is also self-healable.The nociceptor offers noxious signal filtering capability from the sensory information received (tactile, thermal, etc) by eliciting a significant response a) Decentralized computational platform by embedding associative learning elements close to sensor nodes.b) A schematic on how the diffusive memristors are incorporated as nociceptive elements and drift memristors as synaptic elements in a decentralized computational platform.Reproduced with permission. [92]Copyright 2020, Springer Nature.to a high frequency ("noxious") stimulus while low-frequency stimulus generates insufficient response to trigger the nociceptor.Furthermore, intensity of noxious events is reflected in the form of threshold shifts, emulating key nociceptive characteristics as Hyperalgesia and Allodynia.Allodynia can be re-garded as the threshold shift (leftwards for sensitization) in the response-stimulus relation, meaning a smaller signal would be able to trigger the system.While Hyperalgesia is the translational shift (upwards for sensitization), meaning smaller responses are enlarged.This allows for adaptive sensing dependent on the The turning probability that is tuned by the membrane potential.Reproduced with permission. [96]opyright 2021, The American Association for the Advancement of Science.magnitude of external stimulus and if the system undergoes physical damage.Coupled with the self-healing capabilities of the ionic gel, the nociceptor is also able to recover from physical damages inflicted during exposure to noxious stimuli.This selfrepairing ability is crucial for ensuring the reliability of robotic components in real-life applications.The incorporation of elements (see Figure 17b) such as nociception and self-healing capabilities in these decentralized systems, i.e., reflex arc, provides a framework for future intelligent robotics and prostheses.

Sensorimotor Integration with Associative Learning Capabilities
"Muscle memory" or "training makes perfect" is often used as an example of sensorimotor integration in our daily lives.The ability to contain the results of repeated training and influence the outcome of a task can be attributed to the non-volatile memory in the synapse.The strength of the synaptic connection strongly determines the decision-making process (i.e., turning left or right) and can be tuned by training through association.Here, a sensorimotor integration of sensors and actuators in a path-finding robot is illustrated (see Figure 18a). [96]Associative memory is induced through the co-exposure of two stimuli -touch (training) and op-tical (sensing).The robot initially turns right at every intersection, but can be corrected by applying a mechanical touch stimulus that modulates the conductance at the non-volatile memory component (see Figure 18b).As a result, the voltage readout, V M is tuned and the turning probability is adjusted.This mechanical touch stimulus can be regarded as a punishment, and therefore induces reinforcement-type training.

Conclusion and Outlook
In conclusion, this review takes an overview of multi-terminal transistor devices that have advanced beyond simple synaptic emulation.In the pursuit of a synaptic device that could co-localize memory and computing, neuromorphic engineering seeks to reduce the power consumption and device footprint of data-centric applications.Achieving high order reconfigurability in a single device is beneficial in the device footprint point of view.With a smaller footprint with fewer components, significant interference and fabrication issues associated with integration of many individual components can be avoided.This has been shown in a plethora of approaches which could be generalized into two categories.The first focuses on achieving a highly versatile information storage device that can be programmed by many inputs -multiple modes and multiple in-plane electrodes.The second focuses on maximizing the efficiency of transmitting information, in cases of intense noxious stimuli or to efficiently regulate a dynamic response.
However, open challenges such as stability of the emerging materials and large-area integration remain.A few lessons can be derived from this review.Information storage capabilities lie in creating a charge imbalance; the source of imbalanced charges can come from trapped electrons, slow ionic polarizationrelaxation, and electrochemical (physi-sorption, chemi-sorption) processes.Temporal or short-term memory is a result of a kinetic imbalance in charges and higher-order dynamics can be de-mystified as a convolution of more than one source of charge imbalance.To further improve the capabilities and the stability of these devices, the key elements are components of the multiterminal transistor -electrolyte, semiconductor, and stimuli.

Controlling Ionic Movements
Innovations in electrolytes such as controlling ionic transference and introducing mixed ions can be carried out in the future.The electrolytes introduced in this article are generic ionic conductors where both cations and anions are mobile under an applied bias.In some cases, small ionic species such as protons and lithium ions are used to intercalate into the semiconductors without inducing severe crystallographic stresses.In the intercalation systems, a long-lasting charge imbalance can be induced to create long-term memory for learning.Ionic conductors with higher ionic transference can deliver lower energy synapses with enhanced stability.Furthermore, to reduce the stresses from the intercalation processes, a charge imbalance can theoretically be created by electrolyte engineering. [97]Lastly, mixed ions with distinct mobilities can be explored as a source of higher-order temporal dynamics.

Translating Charge Imbalance into Property Changes
High-mobility single layer semiconductors have traditionally be reported in this review article.However, a high field-effect mobility does not necessarily correlate to any figure-of-merit (e.g., endurance, retention) for memory performance.The recent development in heterostructured semiconductors that exhibit unique charge trapping mechanisms can potentially benefit from the charge imbalance to exhibit unique properties.Furthermore, semiconductors that can fully utilize the charge imbalance to modulate electronic or optical properties need to be further explored.This includes strongly correlated material systems such as Mott insulators, niobates, and perovskite nickelates. [98]

Transducers that Create Charge Imbalances
Stimuli modes beyond electrical and light should be explored.An obvious gap in the sensory capabilities can be filled with mechanical (i.e., touch and hearing) and chemical (i.e., olfaction and gustatory) transducers.Perovskite capacitors with piezoelectric (mechanical-to-electrical transduction) or semiconductors constructed from piezoresistive materials can be utilized as multiterminal transistors.Chemiresistive materials utilized in gas sensors and ion electrochemical sensors can also be explored in this domain.In the unconventional computing domain, dynamical properties in materials physics including spintronics, molecular switches [99] and in-materia [100] systems are also expected to further advance this area.
On the second focus that is transmitting information efficiently, the idea is to integrate as much functionalities (information storage and processing capabilities) into the connections as possible.One key enabler could be a device that could co-localize sensing and actuation.By actuation, we are looking at all types of actions beyond mechanical actuation, to changes in optical and chemical transmission properties.Neuromorphic computing is attractive due to its ability to perform memory-intensive tasks more efficiently.A great use case lies in regulating dynamic stimuli.A material or device that could simultaneously sense and apply mechanical forces could regulate a mechanical actuation force on a soft object like an autonomous control system.

Co-Localization of Sensing and Responding
The electrochromic photodetector [101] was an example given in the review article.The device is capable of sensing -that is detecting changes in light intensity.It is also capable of responding -that is controlling the transmission of light through the material.Such a device could potentially be deployed in applications where a dynamic light source needs to be regulated or controlled.

Neuromorphic Control and Regulation
There is a need to expand use cases of neuromorphic or sensorimotor control systems beyond mechanical force regulation.There are several examples of the biological body regulating critical biological functions such as visual attention, breathing, and temperature.With the right development of co-sensingresponding device, neuromorphic control systems have the potential to significantly improve efficiency in signal transmission.Adaptive materials and circuits would also play a critical role for the next-generation bioinspired electronic and robotic applications.
Si En Ng completed his Ph.D. at the School of Materials Science and Engineering at Nanyang Technological University (NTU), Singapore.His thesis focuses on developing reconfigurable photodetectors using stimuli-responsive materials and interfacial barrier engineering.His interests lie in novel materials physics and processing techniques for sensing and information storage.Currently, he is a research fellow working on perovskite-based neuromorphic photodetectors and sensors.
Sujaya Kumar Vishwanath was a research fellow at the School of Materials Science and Engineering at Nanyang Technological University (NTU), Singapore.His pioneering research includes the development of flexible halide perovskite memristors and novel resistive switching devices for neuromorphic computing.He completed his Ph.D. at Kongju National University, South Korea (2017), specializing in transparent conducting oxides and memristors.Currently, he is an Inspire Faculty Fellow at the Indian Institute of Science, India.
Nripan Mathews is an associate professor and the Provost's Chair in Materials Science and Engineering at the School of Materials Science and Engineering at Nanyang Technological University (NTU), Singapore.His interest lies in the development of novel and inexpensive electronic materials through cost-effective techniques for electronics and energy conversion.These include organic-inorganic halide perovskites, metal oxides, and organic thin films.He is interested both in the fundamental of the electronic materials as well as their applications in practical devices such as solar cells, thin-film transistors, and memory devices.

Figure 1 .
Figure 1.Two pivotal approaches in utilizing multi-terminal transistors -coupling inputs by dynamical material properties and enabling efficient data sensing and signal processing.Configurational design and novel materials properties both play critical roles in both approaches.

Figure 2 .
Figure 2. Emulating synaptic plasticity in transistors.a) The transfer characteristics of a transistor.The transistor device comprises a semiconductor channel that is contacted by two source-drain electrodes coupled to a gate electrode.b) The conductance in the transistor is analogous to the strength of the synaptic connections between 2 neurons, programmable with each neuronal inputs.c) A generalization of transistor types based on the semiconductor channel and gate capacitors.

Figure 3 .
Figure 3. Ionic-electronic gate coupling.a) Schematic of PEDOT:PSS array device configuration.b) Local input (I i ) on individual device gates, c) Global input (G) on a common gate, and outputs (O i ) of the individual devices.d) Connection weight as a function of the global gate voltage G for different electrolyte concentrations (DI water-100 mm NaCl).Reproduced with permission.[35]Copyright 2017, Springer Nature.

Figure 4 .
Figure 4. Emulation of homeostatic and hetero-synaptic plasticity via ionic-electronic gate coupling.a) Schematic of dual-gated inorganic synaptic transistor.[38]b) Controlled facilitation and depression with dual gating approach.c) The dual-gating approach augment the plasticity and enable higherorder temporal correlations.Reproduced with permission.[38]Copyright 2018, American Chemical Society.

Figure 6 .
Figure 6.Electronic-photonic coupling.a-c) Triggering of V GS and V DS dependent weight change (Δw) with a pre-synaptic light spike (1.82 mW cm −2 , 100 ms) in In-Zn-O phototransistor.(a) Schematic diagram of an In-Zn-O phototransistor.(b) Triggering of Δw by a pre-synaptic light spike at various V GS values.(c) Δw plotted as function of V DS at various V GS values.[42]d) A 3D schematic illustration of CsPbBr 3 QD-based flash memory.e) Example of photonic potentiation and electrical habituation implementation.Reproduced with permission.[42]Copyright 2018, IOP Publishing.Reproduced with permission.[52]Copyright 2018, John Wiley and Sons.

Figure 9 .
Figure 9. Orientation detection via lateral capacitive coupling in multi-electrodes system.a) Highly interconnected, multi-input and output neural device (left-side) with detailed structure of a single self-supported transistor (right-side) [ b) Schematic of ion-gel gated transistor with two in-plane pre-synaptic gates (x = 0 and x = 1-6) and the corresponding polar diagram of EPSCs with different spatial oriented, input pulse.[70]

Figure 10 .
Figure 10.Hetero-synaptic competition emulated in electrochemical multi-electrodes system.a) Schematic representation of a biological hetero-synaptic junction with limited plasticity-related proteins (PRPs) b) The emulation of a hetero-synaptic junction using a transistor with 4 pre-synaptic terminals.The competition for Li + in the channel is analogous to the limited proteins in biological neural networks.Reproduced with permission.[77]Copyright 2018, Springer Nature.

Figure 11 .
Figure 11.Sensory adaptation with electrochromic photodetectors.a)The shift in sensing threshold with the duration exposed in dark can be emulated with hybrid devices such as the perovskite-MoS 2 phototransistor.The perovskite layer functions as the photosensitizer while the MoS 2 channel is responsible for charge transport.b) The degradation of sensitivity to red light and improvement with blue light when exposed to a train of pulsed stimuli.c) Sensory adaptation to dark as illustrated with the repeated cycles of stimuli and adaptation.Reproduced with permission.[82]Copyright 2022, Elsevier.

Figure 13 .
Figure 13.Adaptive sensory systems based on integrated sensor-memory (1 Sensor + 1 Synapse + 1 Load) circuits.a) Schematic of a light-adjustable optoelectronic neuromorphic circuit analogous to a biomimetic visual perception system.b) A 3×3 array encoded with the letter "H" at both high and low light intensity.The application of appropriate load voltage is analogous to adjusting retina would enhance the discrimination between pixels.Reproduced with permission.[84]Copyright 2019, John Wiley and Sons.c) Schematic of a sensory neuron compared to NeuTap.d) Tactile perceptual learning and pattern recognition by NeuTaP on finger.Schematic of a two-bit binary code for three pattern pairs along with the RI for four groups of training data.e) Error rate recognition as a function of learning times, where the measured data and line fitting are denoted by the black squares and red line, respectively.[85]

Figure 14 .
Figure 14.Perceptual learning systems based on integrated sensor-memory (1 Sensor + 1 Memory).a) Schematic of a sensory neuron compared to NeuTap b) Tactile perceptual learning and pattern recognition by NeuTaP on.Schematic of a two-bit binary code for three pattern pairs along with the RI for four groups of training data.c) Error rate recognition as a function of learning times, where the measured data and line fitting are denoted by the black squares and red line, respectively.Reproduced with permission.[85]Copyright 2018, John Wiley and Sons.

Figure 15 .
Figure 15.Spike generation as a neuronal function.a) Key characteristics of spike generation -threshold, depolarization, and repolarization.b) Organic electrochemical neuron based on Axon Hillock circuit.c) Voltage spike generation with injected current.d) Organic electrochemical synapse based on electropolymerization.e) Synaptic weight facilitation and depression.f)The modulation of the Venus Flytrap using spike-train from the artificial organic electrochemical neuron.Reproduced with permission.[90]Copyright 2022, Springer Nature.

Figure 16 .
Figure 16.Fast-acting detectors based on reflex arc (1 Sensor + 1 Neuron + 1 Actuator).a) An artificial afferent nerve comprising of a pressure sensor, a ring oscillator, and a synaptic transistor b) Discoid cockroach with an artificial afferent nerve attached to its back, hybrid reflex arc made of afferent nerve and a biological afferent nerve along with the photograph of reference and stimulating electrodes, a detached cockroach leg, and a force gauge.Reproduced with permission.[86]Copyright 2018, The American Association for the Advancement of Science c) Decentralized computational platform by embedding associative learning elements close to sensor nodes.d) A schematic on how the diffusive memristors are incorporated as nociceptive elements and drift memristors as synaptic elements in a decentralized computational platform.[92]

Figure 17 .
Figure 17.Fast-acting detectors with learning functionalities (1 Sensor + 1 Nociceptor + 1 Actuator for reflex arc, 1 Neuron + 1 Synapse for learning).a)Decentralized computational platform by embedding associative learning elements close to sensor nodes.b) A schematic on how the diffusive memristors are incorporated as nociceptive elements and drift memristors as synaptic elements in a decentralized computational platform.Reproduced with permission.[92]Copyright 2020, Springer Nature.

Figure 18 .
Figure 18.Sensorimotor integration with associative learning capabilities.a) Path-finding robot navigating through sensor signals processed by organic neuromorphic circuit capable of associative learning, where the actuators are controlled autonomously.b) Mixed digital and analog processing.Analog component involving a voltage divider circuit that contains the non-volatile synaptic conductance and the training outcome is converted into digital signals for the processor that controls the actuators.c)The turning probability that is tuned by the membrane potential.Reproduced with permission.[96]Copyright 2021, The American Association for the Advancement of Science.