Recent Progress in Synaptic Devices Paving the Way toward an Artificial Cogni‐Retina for Bionic and Machine Vision

The state‐of‐the‐art conception of a bionic/robotic eye is a somewhat bulky multipart system comprising a video camera connected to a processing unit that in turn communicates data through a wireless transmitter to either an in vivo retinal implant or a computer system. An artificial cogni‐retina is a millimeter‐scale, intelligent apparatus designed as a replacement for these systems, while executing simple image processing tasks. As a bionic limb, it can connect directly to the optic nerve and perform rudimentary cognitive functions such as perceiving, learning, remembering, and classifying elementary visual data. This theoretical system presents a quantum leap in terms of size, power consumption, and speed to both prosthetic human eyes and robotic vision in artificial intelligence‐based platforms such as autonomous vehicles. Recently, an increasing number of publications have used interesting materials in artificial synaptic devices that drive this idea closer toward becoming a real‐world application. Such devices may form a basis for hardware‐based deep learning artificial neural networks that can potentially execute image processing tasks within a single clock cycle compared to software algorithms running on conventional von Neumann machines that require millions of cycles to perform image sensor interfacing, memory fetch operations, and data path propagation.


Introduction
The research on brain-inspired computing algorithms and artificial neural networks (ANNs), aimed at mimicking the way the brain processes data and learns, has been ongoing since the introduction of the "learning machine" concept by Alan Turing in 1950. [1,2]Some important milestones in this field are the construction of the stochastic neural analog reinforcement calculator (SNARC) [3] and the invention of the Perceptron by Frank Rosenblatt. [4,5]Within this realm, deep neural networks are computational methods inspired by biological nervous systems for the extraction of general relations from big unstructured and unlabeled data sets.Such ANNs have proven to be extremely efficient in performing generic tasks such as pattern recognition. [6,7]owever, they are almost exclusively implemented as software algorithms running on von Neumann machines.The serial nature of logic-memory processing in these machines is fundamentally incompatible with the massive parallelism associated with intertwined neural networks.This incompatibility presents a serious drawback in terms of processing time and power consumption, especially for applications (e.g., autonomous vehicles and natural language processing).
10][11] A typical neuron can have several thousands of synapses (%10 4 ) that mostly connect axons in a presynaptic neuron to dendrites in postsynaptic neuron.Interneural signaling occurs by the release of neurotransmitters from the presynaptic neuron into the gap (i.e., synaptic cleft) that in turn are collected by receptors in the postsynaptic neuron.Learning processes in the brain occur by strengthening and weakening of synaptic connections throughout neural networks.This process is commonly termed as functional plasticity.In this context, short-term plasticity (STP) is a term used to describe temporary rapid changes in the synaptic strength that can occur on a timescale from tens of milliseconds to a few minutes.Long-term plasticity, on the other hand, can occur from minutes to hours.In addition, long-term potentiation (LTP), which is the strengthening of the synaptic weight over time, is related to memory storage, whereas long-term depression (LTD) weakens the connection and clears unwanted memories and synaptic paths.A combination of LTP and LTD triggered by spike-time correlations is called spike-time-dependent plasticity (STDP).The postsynaptic response is classified as being either excitatory or inhibitory and determined by the type of neurotransmitter used during signaling.Artificial synaptic devices (ASDs) aim to mimic as much as possible the features that biological synapses have and in particular, the ability to adjust the weight parameter as a function of activity.
Memristors [12] (devices with reversible analog resistance) have been demonstrated as being capable to successfully emulate the characteristics of a biological synapse.From a hardwareimplementation point of view, memristor crossbar arrays configured as ANNs have gained momentum in the past decade after the demonstration of the electric field-modulated reversible resistive switching effect (attributable to defect creation/annihilation by ionic motion within a dielectric material sandwiched between two metal electrodes). [13]These architectures are considered as promising candidates for achieving ultra-high densities (%10 15 bits cm À2 ), similar to the human cerebral cortex (%10 15 synapses), especially when configured in a 3D stacking. [14]n addition, they attracted much interest due to their inherent structure that presents a somewhat obvious transition toward the implementation of the hardware-based equivalent to deep learning algorithms [15] running on traditional von Neumann machines. [16]Application-specific neuromorphic computers are specifically tailored to execute task with orders of magnitude of improvement in speed and power consumption in comparison.
The resistive switching effect was marked early on as a promising candidate for artificial synaptic functionality.After the initial booming in demonstrations of this effect over transition metal oxides, [17] silicon oxides, and electrochemical metallization, [18] the next evolutional step in ASD realization, from a material science point of view, was the incorporation of more intriguing active layers to mimic the corresponding biological functionality.Some implementations were demonstrated using organic-inorganic hybrid halide perovskites commonly used in optoelectronic applications such as solar cells, photodetectors, and light-emitting diodes. [19]Others relied on polymer-based, [20] magnetic, [21][22][23] and organic electronic [24][25][26][27][28][29][30] materials to achieve the resistive switching effect.More recently, an increasing number of publications have presented the use of interesting materials that show a potential for innovative platforms beyond the mainstream concepts.These works present a step toward the implementation of futuristic platforms such as bionic "thinking" retinas.The publications focus on interesting material combinations and associated properties range from low dimensional, flexible, biocompatible, and photonic electronics hybrid responsive.The current report presents these developments in light of their prospective relevance to this application.This report summarizes some of the most recent important developments in devices for neuromorphic computing in this context and how these may be used for implementing the goal of a compact and efficient artificial cogni-retina.In addition, current limitations and future research requirements are discussed.The report is divided into sections based on material classifications and associated properties, in which the current limitation and future requirements are also discussed.

Artificial Cogni-Retina
An artificial cogni-retina would be able to capture visual images and perform rudimentary image processing tasks.A bionic system would convey the data directly to the optic nerve as shown in Figure 1.Ideally of millimeter size and without external modules (directly implantable into a human eye), such a prosthetic system would be superior to state-of-the-art bionic eyes (Figure 1a) in almost every aspect.A novel approach for implementing this system makes use of deep-learning hardware-based ANNs.Recent progress in this area has led to ASDs having a wide range of properties and functionalities.The main advantage of this approach lies in the construction of queued arrays (Figure 1b) capable of performing extremely fast, analog computations that correspond to mathematic matrix operations.[33][34][35][36] By using the same concept while forfeiting a biocompatibility restriction, the cogni-retina could function as a robotic eye.
The first layer (highlighted in red) in a series of ASD-based crossbar arrays that make up a cogni-retinal apparatus is shown in Figure 1b,d,e.The primary function of this layer is to act as photodetector that captures visual information by imprinting it over an array of ASDs.Each pixel data correlates to the bitwise synaptic weight of an individual device located in a corresponding placement in the array.Once the image is captured, in-memory processing of the entire image data can occur through the successive layers.Crossbar arrays are extremely efficient, in terms of calculation time and energy, when performing matrix dot products for deep learning algorithms.For this purpose, a prospective building block device must be able to respond to both photonic and electric stimulations that alter its synaptic state (LTP and LTD).successive layers in a deep learning neural network.As intended to operate in vivo, these arrays should be based on biocompatible materials that can be safely incorporated into living tissue.The main purpose of this construction is to reduce the complexity of the perceived data by taking in some of the functions usually associated with the visual cortex.In addition, it decreases the Figure 1.a) Depiction of a bionic eye system.Visual information is captured by a glass-mounted camera and passed to an image processing unit that in turn processes and transmits it.A receiver system placed on a human eye then generates electrical signals that stimulate the retina via an array of implanted electrodes.b) A depiction of the bionic cogni-retina concept constructed as a deep ANN having successive 3D crossbar memristor array layers (red and blue) and an output layer connected directly to the optic nerve (green).Each layer is based on ASDs that allow the determination of the synaptic weights according to different designated functionalities.The first layer translates a visual image to an imprint of bitwise synaptic weights onto the first crossbar array (red).The following layers perform basic image processing computation by using predetermined synaptic weight values (blue).Eventually, an output layer translates electronic signals to chemical ones and communicates the processed visual information directly into the optic nerve (green).c) Visualization of the placement of a cogni-retinal implant inside the human eye, which presents a much more compact and convenient prosthetic option.d) A block diagram showing the concept of a deep ANN having cascaded layers based on dedicated ASD.The input layer (red) is used to capture visual information and is based on hybrid photonic-electronic ASDs.Two hidden layers (blue) execute rudimentary image processing functions using biocompatible ASDs.Finally, an output layer (green) is used to translate the electronic signals to chemical data directly into the optic nerve through organic devices.e) Perception flow of a shape (square) according to the connections and arrangement of the successive arrays.
number of required subsequent connections to the optic nerve.The number of cascaded arrays corresponds to the number of hidden layers in a deep learning algorithm.The designated output layer (highlighted in green in Figure 1b,d,e) conveys the computation results directly to the optic nerve.It is based on electric to chemical signal translation organic devices.
The perception flow of a basic shape (square) is depicted in Figure 1e (symbolic diagram on the upper right) and demonstrates how the cogni-retina not only captures visual data but also performs some basic analysis usually associated with the visual cortex.The shape is imprinted on the first input layer (red) as a photoresponse.This in turn triggers an electric path of predetermined synaptic weights (blue arrows) that eventually invoke a "square identified" signal at the output (green arrow).This output drives a section of the optic nerve and passes the data to the visual cortex.In this manner, some image processing is performed already at the source (retina) and the number of required connections to the optic nerve is reduced significantly.

Hybrid Photonic Electronic Responsive ASDs
Several publications in the previous year introduced photoresponsive material-based constructions in ASDs.These devices present a feasible option for the implementation of light-responsive ANNs.Wang et al. [37] introduced a flexible, light-triggered organic neuromorphic device.The device can translate incident light signals of different intensities and wavelengths into volatile or nonvolatile electrical signals.It is based on a ferroelectric/electrochemically modulated organic structure and capable of mimicking the operational mechanism of biological synapses such as STP, electrochemical dopinginduced LTP, and ferroelectric switching-induced LTP.A gate bias that yields an electric field below the electrochemical doping limit (Ec) produces an STP signal.Repetitive presynaptic spikes induce electrochemical LTP behavior, while field levels exceeding Ec switch the dipole orientation to trigger ferroelectric LTP.
The device presents an integration of a synaptic device and a light-sensitive element, which emulates the visual perception and converts incident photons into synaptic signals.Figure 2a depicts rods and cones interacting with visible light to produce signals that are transmitted through synapses to bipolar and ganglion cells.These signals in turn generate action potentials that are communicated to the visual cortex section in the brain through the optic nerve.The 3D structure of the device is shown in Figure 2b.It is based on an organic transistor gated by P(VDF-TrFE)/P(VP-EDMAEMAES) dielectric.Figure 2c shows the mechanism of STP, electrochemical doping-induced LTP, and ferroelectric switching-induced LTP in the synaptic transistor.65% of the initial level.Reproduced with permission. [37]Copyright 2018, Wiley-VCH.
When the gate potential induces an electric field (Ec) below the polarization P(VDF-TrFE), the electrochemical doping produces STP signals.Repetitive presynaptic spikes induce electrochemical LTP.When the field exceeds Ec, the dipole orientations flip and trigger ferroelectric LTP.
The authors were able to demonstrate wavelength-recognition functionality by irradiation of light with identical intensities (10.80 mW cm À2 ) with different frequencies (850 and 550 nm) on an array of devices and showed near infra-red (NIR) triggered electrochemical LTP.Using NIR illumination, signals in the exposed area ranged from 1.0 to 1.7 nA (Figure 2d) and after 600 s, the current decayed by an order of magnitude (Figure 2e).After 1800 s, the irradiated section was virtually distinguishable (Figure 2f).Alternatively, the 550 nm green light irradiation resulted in higher current levels (0.40-0.67 μA in Figure 2g).After 600 s of illumination, about 78% of the signals were still intact (Figure 2h).However, the intensity was reduced to 65% after 1800 s (Figure 2i).
Zhu and Lu [38] demonstrated a CH 3 NH 3 PbI 3 (MAPbI 3 )-based memristor that exhibits light-tunable synaptic behaviors.The conductance of the device was modulated in real time by illumination events.It was shown that by increasing the formation energy of iodine vacancy, light illumination can be used to control their generation and annihilation dynamics in a similar manner to the flux of Ca ions in biological systems.In this manner, the device was progressively switched back to a high resistive state (HRS) after forming a low resistive state (LRS).The dynamics affected by illumination intensity is shown in Figure 3.The illumination intensity and wavelength were demonstrated to control the formation and reset of the memristive devices with fast response times (≤20 ms).The conductance was shown to drop by 20% of its original level of 25 μS with an illumination intensity of 0.1 μW cm À2 .
The voltage required to turn the device into an LRS (V 1 ) increased under illumination with direct relation to the intensity (as indicated by triangles in the current-voltage (IÀV) curves of Figure 3b).In addition, the potential needed for resetting it back to an HRS (V 2 ) increased (as indicated by stars in the IÀV curves of Figure 3b).Basically, light radiation made the formation more difficult and the resetting easier.This confirmed that irradiating the device's surface both inhibited iodine ion and vacancy formation and accelerated their recombination.The removal of the electrical bias under light illumination also accelerated the conductance decay process.
He et al. [39] demonstrated photonic potentiation and electric habituation in ultrathin molybdenum disulfide (MoS 2 )-based memristive devices.The authors could mimic synaptic neuromorphic functions, such as potentiation/habituation and short-term/long-term plasticity based on inherent photoconductivity and a volatile resistive state due to spontaneous trapping/ detrapping processes.Figure 4 shows an overview of their experimental results using I-V sweeps under UV light (310 nm) illumination and in darkness.In darkness, the p-n junction showed near perfect rectification, with a large self-rectification ratio of about 4 Â 10 3 under a bias of 2 V (black curve in Figure 4a).The temporal evolution of the photocurrent in the ambient atmosphere is shown in Figure 4b.A fast current response (stage 2) was attributed to the band-to-band transition that creates excited electrons and holes.After this initial response, the current continued to increase gradually (stage 3) to an order of magnitude above the dark level.The authors attributed this phenomenon to the trapped photo-generated holes.Photonic potentiation and electric habituation were realized by applications of consecutive light and electrical pulses, respectively, as shown in Figure 4e.The current was initially at a level of 5.6 nA.After the application of a series of light pulses (0.11 mW cm À2 for 1 s), the conductance increased gradually due to the generation of electron-hole pairs.In the next stage, a series of negative electric pulses were applied (À8 V for 5 ms).The electrons in the MoS 2 layer were driven to the interface, captured by interfacial trap sites, and the conductance gradually decreased.
An intriguing class of materials that attracted much attention in the context of combined photo-electro functionality is perovskites commonly for solar cell devices.As these are emerging in the implementation of artificial synapses, [19,40,41] it would be worth covering some of the work done in this field that might be used as a future implementation of a hybrid-responsive ASD.Xiao et al. [42] demonstrated a large switchable photovoltaic effect in low-cost organometal trihalide perovskite-based structure (indium-tin-oxide [ITO])/poly(3,4-ethylenedioxythiophene): poly(4-styrenesulfonate) (PEDOT:PSS)/perovskite/Au. The authors fabricated photovoltaic devices with vertical and lateral cell configurations and observed that the photocurrent direction can be switched repeatedly by applying a small electric field of less than 1 V μm À1 .The switchable photocurrent reached a level of 20.1 mA cm À2 using one sun illumination condition.
Nie et al. [43] reported that the slow photocurrent degradation in thin-film large-grain perovskite-based photovoltaic devices is due to the formation of light-activated meta-stable deep-level trap states.However, the devices can self-heal completely by resting them in the dark.In addition, the degradation can be completely prevented by operating the devices at 0 C.They attribute the observed photocurrent degradation to the formation of lightactivated meta-stable trap states, which over prolonged illumination leads to the accumulation of space charges showing self-healing of less than 1 min at 70 C.The self-healing phenomenon is interesting by itself because it can relate to certain aspects of biological nociceptors.The latter are sensory neurons that respond to potentially damaging stimuli and exhibit special characteristics such as threshold, self-healing, hyperalgesia, and allodynia.Artificial nociceptors have been demonstrated to mimic this behavior. [44,45]alide perovskites (ABX 3 ) are alloys that have been widely studied with respect to bandgap engineering in the visible range and beyond.As such, they present an interesting option for implementations of wavelength-dependent photonic responsive ASDs.In this context, Kim et al. [46] reported that light excitation enhances the ionic conductivity of methylammonium lead iodide by several orders of magnitude, resulting in a large tunable photo effect on ion conduction.Lei et al. [47] could tune the bandgap of single crystalline CsPb x Sn 1Àx I 3 nanowires from 1.3 eV to 1.78 eV by varying the Pb/Sn ratio, which leads to tunable photoluminescence in the NIR range.In addition, they found that the electrical conductivity increased when more Sn 2þ was alloyed with Pb 2þ , possibly due to the increase in charge carrier concentration.
Recent works conducted on black phosphorus have also shown a feasible possibility in this respect.Tian et al. [48] fabricated a three-terminal modulated nonvolatile memory devices using this material.They showed dynamically reconfigurable and polarity reversible behavior as a function of gate bias by switching the polarity of carriers from electrons to holes.Electron-and hole-dominated conductions were defined as the erased and programmed memory states.Whitney et al. [49]  ) was alternately applied.k) Evolution of the conductance when stimulated with electrical pulses (1 V, 10 ms, 1 ms interval), in the dark and under light illumination (1.29 μW cm À2 ).Reproduced with permission. [38]Copyright 2018, American Chemical Society.modulated the optoelectronic response of quantum-confined carriers in black phosphorus using the field effect.The optical response also depended on flake thickness and carrier concentration.By measuring the transmission and extinction amplitudes, a modulation of about 2% was shown.Both research results are summarized in Figure 5.
Figure 5a shows the structure of the reconfigurable charge trap memory device [48] using black phosphorus as the channel material.This device can provide ambipolar hole and electron conduction modes with different back-gate voltages (V BG ).The thickness of the channel material is about 10 nm as shown in the atomic force microscope image of Figure 5b.The small  [39] Copyright 2018, Wiley-VCH.
bandgap of this material, along with the used metal contacts, allows for both electron and hole injection into the channel.A positive V BG can tune the Fermi level close to the conduction band edge (n-type), while a negative bias can tune it close to the valence band edge (p-type).As a result, the device is in the OFF state for V BG ranging from À7.92 to þ6.24 V as seen in Figure 5c.The transfer curve of the memory device for both forward and backward top-gate voltage sweeps at V BG ¼ 0 V is shown in Figure 5d.The two hysteresis windows indicate the ambipolar transport nature of the device.
Whitney et al. [49] utilized the infrared optical properties of black phosphorus that result primarily from the unique band structure of a 7-nm flake (Figure 5e).Sub-bands arising from vertical confinement allow for quantized inter-sub-band transitions and provide the primary contribution to its zero-field optical conductivity.
Their primary results showed the modulated extinction of the sample under different applied voltages when normalized to the zero-bias extinction spectrum (Figure 5f).The spectra showed three distinct features.First, under negative applied bias (i.e., when the sample is being depleted of holes), a negative peak (I) appears in transmittance near 0.45 eV, which grows in amplitude and broadens to lower energies as the magnitude of the bias increases.Second, under positive applied bias (i.e., when the sample is being increasingly hole-doped), a positive peak (II) appears in transmittance near 0.5À0.7 eV.Finally, these two effects are superimposed with an oscillatory feature (III) that varies with the magnitude of the applied field, but not with its polarity, which is most clearly visible in the negative bias spectra in the 0.5À0.7 eV range.
Although significant progress has been made recently in demonstrating the proof of concept for hybrid f) Fourier-transform infrared spectroscopy (FTIR) transmission extinction versus photon energy normalized to zero bias.a-d) Reproduced with permission. [48]Copyright 2016, American Chemical Society.e,f) Reproduced with permission. [49]Copyright 2016, American Chemical Society.
photo-electro-responsive ASDs, a realistic implementation of a visual detector mandates more improvements, especially in device sizing.Resolution is a key feature in the realization of the first layer of a deep neural network cogni-retina application.As mentioned previously, the purpose of this layer is to capture visual data by imprinting it over an array of photonic-responsive ASDs.This in turn dictates the need for highly integrable, building-block devices.Ideally, these devices would be based on a twodimensional (2D) active layer material, configured in a lateral manner with the electrodes placed vertically.Other important key features are low power dissipation, operating current, and voltages.For an application that is intended to function as an in vivo prosthetic implant, power consumption that determines heat production is a crucial feature.Low-current functionality may allow future electric triggering by biological nerve impulses.
Nano-scale resistive memory devices for neuromorphic computing were already demonstrated to have femto-Joule energy consumption, [50][51][52] and however, not photo-responsive.The devices discussed in this section are configured laterally in the micrometer regime and still fall short when measured by these criteria.
Optical assisted recovery of soft breakdown (SBD)-induced defects in metal-oxide thin-film materials (ZrO x and HfO x ) were observed by Zhou et al. [53] In the context of this section, these materials presents an interesting option for the implementation of ASDs, as these materials have been extensively used to demonstrate resistive switching effect-based synaptic behaviors. [54]igure 6a,b shows a schematic illustration of the photo-assisted interstitial-vacancy recombination for negative photoconductivity response.After soft electrical breakdown, a conductive path Reproduced with permission. [53]Copyright 2018, AIP Publishing.comprising a cluster of oxygen vacancy defects is formed.Some of the oxygen ions from the dissociated metal-oxygen bonds propagated outward and are situated at interstitial sites around the conductive path.During illumination, subsequent recombination with the vacancy defects results in the interruption of the conductive path and decrease in the breakdown leakage current.The authors could link the recovery of the SBD-induced vacancies in the film with white light illumination as shown in Figure 6c-f.The LRS was thus reset back to an HRS upon irradiation of light.
In addition to the above, Yang et al. [55] fabricated photoelectric neuromorphic devices based on pulse light-stimulated lowvoltage indium-gallium-zinc-oxide (IGZO) electric double-layer transistors.The device was based on a 2D structure IGZO channel over a chitosan electrolyte layer controlled by an ITO bottom gate.Protons driven by the back gate onto the channel interface were used to modulate the channel conduction of electrons that in turn were excited by light irradiation.Their photoelectric neuromorphic devices were shown to mimic some synaptic functionality, such as long-term plasticity and transition from depression to potentiation by gate voltage modulation.
Some publications aimed to duplicate the behavior of complex synaptic interaction associated with spatiotemporal perception.Xie et al. [53,56] demonstrated a facile coplanar multi-gate 2D MoS 2 electric-double-layer transistor with proton-conducting poly(vinylalcohol) electrolytes as laterally coupled gate dielectrics that mimicked visual neurons in the visual cortex.Their device showed excitatory postsynaptic current (EPSC) and paired-pulse facilitation (PPF).Their structure aimed to mimic the way that the retina conveys information through the thalamus to the visual cortex by using a multi-gate (nine separate gates laterally structured next to an MoS 2 channel) device.In this manner, they demonstrated orientation recognition in the visual system.Although the device did demonstrate a complex, orientation-dependent excitation and inhibition characteristics, it was not, and cannot be classified as light responsive.

Biocompatible ASDs
A biocompatible, flexible, and transparent resistive switching memory device based on a naturally abundant organic polymer material was demonstrated to show resistive switching behavior in 2015 by Hosseini et al. [57] The device had a coplanar structure of Mg/Ag-doped chitosan/Mg on plastic substrate as depicted in Figure 7b.An observed bipolar resistive switching behavior was attributed to trap-related space-charge-limited conduction in high resistance state and filamentary conduction in low resistance state.In addition, low power operation and long data retention were demonstrated on the devices when fabricated on natural substrates such as chitosan and rice paper for low-cost applications. [58]The I-V curves in Figure 7c show the setting of the device to LRS (about 1.5 V) and resetting to HRS (about À1 V). Figure 7d depicts the results for a data retention test with a bias of 0.14 V under ambient condition.The device showed an ON/OFF ratio of 10 2 with little deterioration for over 10 4 s. Figure 7e depicts a reliability test performed for 60 cycles that showed consistent endurance.The cumulative distribution of the set and reset voltages is shown in Figure 7f.Chitosan-based device was also demonstrated by Yu et al. [59] to design a polysaccharide-gated flexible ITO synaptic transistor.
Artificial synapses with short-and long-term memory based on renewable materials were demonstrated by Park et al. [60] They fabricated flexible bio-memristor devices based on lignin, which is one of the most abundant organic polymers.Apart from being flexible, this material is biocompatible, biodegradable, and environmentally friendly.The fabricated ASD was shown to perform memory switching, STP, long-term plasticity, spike-ratedependent plasticity, and short-term to long-term transition.They used Au (an inert metal) as the top electrode (TE) along with an ITO bottom electrode (BE) to minimize the effect of diffusing metal ions on the function of the device.The TE emulated presynaptic neurons, and the BE emulated postsynaptic neurons.Electrical pulses that emulate synaptic spiking events were applied to induce the synaptic behaviors in the as depicted in Figure 8.
This artificial synaptic behavior is depicted in Figure 8a.When five consecutive sweeps of negative DC voltage sweeps (0 to À0.7 V) were applied, the current level increased following each sweep in a similar manner to synaptic potentiation.When followed by five consecutive positive bias sweeps (0 to þ0.7 V), the current level decreased after each sweep in a similar manner to synaptic depression.The conductance at 0.1 V increased or decreased gradually for 10 consecutive sweeps (Figure 8b).The current responses as a function of time and bias levels are shown in Figure 8c.It was shown that gradual changes in conductance can emulate a successively variable synaptic weight or connection strength.The authors demonstrated the conductance modulation over a larger scale of successive pulsation scheme of 50 identical negative pulses (À0.7 V, 100 ms) followed by 50 positive pulses (þ0.7 V, 100 ms).The state of the device was estimated using a small read pulse, having an amplitude of 0.1 V, which was applied after each pulse.These results are summarized in Figure 8d.
Dang et al. [61] used a different approach where they implemented a degradable biomimetic synaptic device based on a W/MgO/ZnO/Mo memristor with a silk protein substrate.Their devices were dissolved completely in phosphate-buffered saline solution or deionized water within 7 min.These devices demonstrated high stability and uniformity of electrical properties as indicated by 300 cycles of I-V sweeping as shown in Figure 8f.The endurance properties of the W/MgO/ZnO/Mo memristors along with the HRS and LRS (measured at 0.1 V) are shown in Figure 8g.Data retention characteristics were tested at room temperature for 10 4 s with a read voltage of 0.1 V, as shown in Figure 8h.The HRS was stable at around 8 Â 10 3 Ω, and the LRS at about 1.05 Â 10 3 Ω.Moreover, the authors reported the possibility of implementing multilevel résistance using these devices.
Kim et al. [62] produced biocompatible artificial synapse based on a matrix of the biodegradable and ecologically benign ι-carrageenan (ι-car) biopolymer.By exploiting Ag dynamics, the authors demonstrated both STP and LTP in their device.In addition, this artificial synapse was shown to emulate PPF and transition from STP to LTP of a biological synapse.These characteristics were realized by exploiting the similarities between the Ag dynamics within the ι-car matrix and the Ca 2þ dynamics in a biological synapse.In addition, the authors showed that an array of 5 Â 5 ASDs was capable of memorizing the letters "P" and "T".
Lee et al. [63] fabricated a biocompatible and transparent synaptic device based on collagen extracted from fish skin and Mg electrodes on ITO-coated polyethylene terephthalate (PET) flexible substrates.An interesting feature demonstrated by this device was mechanical flexibility of the electronic synapse when bent under tensile stress with a 7 mm radius of curvature (Figure 8i,j).Compared to other devices covered in this section, it showed asymmetric I-V characteristics in the bent state that could allow the incorporation of the physical state as a parameter into the ANN computation scheme.
Liu et al. [64] fabricated laterally proton-coupled transistors on free-standing chitosan membranes that displayed synaptic functionality.As chitosan is a linear biopolysaccharide composed of randomly distributed b-(1-4)-linked d-glucosamine and N-acetyl-d-glucosamine (acetylated unit), it is used in wearable electronics due to its flexibility.The authors demonstrated freestanding synaptic devices with multiple presynaptic inputs and spiking logic operation and logic modulation.Their devices have a potential in a range of applications, such as bendable smart tags and synaptic skin-attached electronics.
When reviewed in the light of a bionic retina perspective, each of the biocompatible devices reviewed in this section shows interesting features that address some of the required attributes.However, no single device may be labeled as perfect in this light.Inorganic-based active layer materials placed on an organic substrate [61] allow for an order of magnitude reduction in physical dimensions (although this might just be dictated by available resources and much smaller sizes may be achieved using advanced fabrication techniques) in comparison to the organic polymer-based active layer devices.On the other hand, these devices had an operating current that was in the order of magnitude larger than the bio-active layers, along with relatively higher programming voltage.Flexibility and transparency are key features for inter-retinal implants.A physically flexible device  [57] Copyright 2015, Wiley-VCH.may be used to fabricate arrays that conform to the natural curvature of the internal eye structure.In addition, transparency allows for more freedom in the placement of the functional layers depicted in Figure 1b (i.e., the photo-responsive layer does not have to be placed on top).Possible candidates to serve as ANN building blocks were presented. [60,63]Unfortunately, the devices were relatively large and sized in the millimeter regime.Advanced image processing calculations have a direct relation to  [60] Copyright 2017, American Chemical Society.e-h) Reproduced with permission. [61]opyright 2018, Royal Society of Chemistry.i,j) Reproduced with permission. [63]Copyright 2018, Wiley-VCH.
the overall number of devices in the ANN (resolution) and thus dictate large densities.

Electric to Chemical Signal Translation Devices
The design of artificial devices intended for direct interfacing with the nervous system face major challenges.The most dominant one is probably realizing the ability to translate between the signal carriers of the nervous system (ions and neurotransmitters) and those of conventional electronics (electrons).Neural signals, usually referred to as action potentials, induce the flux of sodium and potassium ions through the neuronal membranes.These changes the concentration of the electric potential in the extracellular fluid.One interesting approach used for the implementation of such devices is the organic electronic ion pump (OEIP). [65]This concept relies on the use of organic conjugated polymers that utilize both electrons and ions as charge carriers.This OEIP is made of the polymer-polyelectrolyte system PEDOT:PSS and translates electronic signals into electrophoretic migration of ions and neurotransmitters.It can operate as a "machine-to-brain" interface by modulating brainstem responses in vivo.
The planar device depicted in Figure 9a was implemented by Tybrandt et al., [66] and they demonstrated OEIP transportation of the neurotransmitter acetylcholine (ACh), which is a major neurotransmitter of the nervous system.ACh is a quaternary ammonium compound that maintains a positive charge across the full pH range.The device was designed with a 10-μm-wide delivery channel (Figure 9a) with a waste electrode to improve the temporal control of delivery.To demonstrate conjugated polymer devices as a communication interface between electronic components and the nervous system, the authors showed that the transported compound retained its biological activity after being transported through the polymer.To analyze the biological activity of the transported ACh, human SH-SY5Y neuroblastoma cells (known to express the AChRs) were used as biosensors (Figure 9a).In this manner, by recording the intracellular Ca 2þ response in cells located next to the outlet point of the 10-μm delivery channel, it was confirmed that ACh retains its biological activity after transport.
The channel was pre-filled by applying a potential between the source and the waste electrodes (V S-W ¼ 20 V) for 10 min.In the active mode (V EC ¼ 10 and V EB ¼ 4 V), the base supplies the junction with Cl À .Increased conductivity results in ACh transport from emitter to collector.In the off-mode (V EC ¼ 10 V and V EB ¼ À1 V), the base depletes the junction of Cl À , and ACh delivery stops due to decreased conductivity.d) Intracellular Ca 2þ recording of ACh-stimulated SH-SY5Y cells cultured on the collector terminal.Turning the base on/off regulates ACh delivery.a,b) Reproduced with permission. [66]Copyright 2009, Wiley-VCH.c) Reproduced with permission. [68]opyright 2011, American Chemical Society.d) Reproduced with permission. [67]Copyright 2010, National Academy of Sciences.
Delivery of ACh to the cells was then initiated by applying a potential between the source and target electrodes.When the potential is turned off, ACh delivery stopped.Due to the diffusion of ACh into the electrolyte, the high local concentration at the delivery point rapidly decreased.Utilizing this effect of controlled delivery versus diffusion, the authors demonstrated the generation of an amplitude and frequency modulation (AM and FM) output signal.The AM signaling pattern in cells could be controlled by increasing the applied voltage for a given pulse length, resulting in an increased delivery rate of Ach, and an increased amplitude of the cellular Ca 2þ response was observed for each extended pulse length. [66]ncreasing the pulse length for a constant voltage can increase the amplitude of the Ca 2þ response to effectively generate an FM signaling pattern as shown in Figure 9b.The temporal plot shows that after extending the duration (from 0.2 to 2 s), the amplitude of the Ca 2þ responses in the cell located at 50 μm from the channel outlet is enhanced.Moreover, while a Ca 2þ response is triggered in the cell located closest to the channel outlet by ACh delivered when applying a 20 V pulse for 0.2 s, the cell located at 150 μm from the outlet only responded when the pulse length was extended to 2 s.In this manner, an OEIP device can be used for modulation of ACh delivery to single cells with a spatial resolution of 100 μm (Figure 9b, inset).The OEIP can also be used to generate temporal patterns that mimic naturally occurring Ca 2þ oscillations in cells by matching the length and strength of the applied pulse of ACh delivery with a periodicity of about 100 s.
The authors later extended their work to demonstrate an electrophoretic chemical transistor with signal amplification, the ion bipolar junction transistor (IBJT) [67] followed up by an anionic equivalent, an npn-IBJT, [68] controlling the delivery of negatively charged ions and integrated chemical logic gates. [69]The IBJT consists of three channels (emitter, collector, and base), which are connected to terminals of PEDOT:PSS as shown in Figure 9c.The cation-selective emitter and collector channels consist of over-oxidized PEDOT:PSS, while the base channel is made of an anion selective membrane.The channel junction interface is based on a neutral cross-linked conductive gel layer.Transport of ACh from the emitter to the collector requires that forward bias is applied across the emitter-collector (V EC > 0 V).In the active mode (V EB > 0 V), chloride ions (Cl À ) migrate through the base into the junction where they are compensated by cations (ACh) from the emitter.This increases the ionic conductivity between the emitter and collector, thereby allowing positively charged ACh molecules to be transported from the emitter, through the gel, and to the collector as a function of the base potential.The temporal behavior of the device is shown in Figure 9d.By reversing the biased voltage of the base, the device is switched on.As Cl À ions move through the base channel into the junction, it enables the motion of ACh from emitter to collector.The delivery of ACh to cells located at the outlet on the collector electrode was thus demonstrated by monitoring the cellular Ca 2þ response. [67]When the IBJT is turned off, Cl À ion migrates from the junction, back to the base, and ACh delivery is terminated.In this context, a tube-like, flexible device was designed by Simon et al., [70] as a more suitable OEIP operating as a machine-to-brain interface.This tube-like shape has potential to be surgically implanted and placed directly in vivo, where the tip from the source and target electrodes (outlet/inlet) comes in direct contact with the target system (any type of biofluid).
The subject of advanced materials in neural interface engineering and implanted neuroprosthetic and biomaterials systems has been reviewed recently in great detail, published in a special issue of Advanced Functional Materials. [71]As part of this issue, the review by Wellman et al. [72] discusses the key issues and recent approaches for handling the biocompatibility aspects of neural probes because the biological inflammatory response to implants, neural degeneration, and long-term material stability diminish the quality of the interface over time.In this regard, recent engineering efforts have focused on optimizing individual probe design parameters such as softness or flexibility.In addition, Fairfield [73] and Ferro et al. [74] reviewed new technologies in electronic and ionic materials for neurointerfaces that form the basis for communication between living brain tissue and artificial devices.New materials that result in lower tissue damage, reduced immunogenicity, long-term stability, and high sensitivity are being integrated to smaller dimensions with the ultimate goal of producing a bidirectional, single cell interface.

Low-Dimensional ASDs for Robotic Vision
This section is intended to cover some of the research focused on the use of 2D materials in the implementation of synaptic devices.An ANN targeted at advanced image processing capabilities would dictate the need for larger scale networks with many deep layers.Using the same concept while forfeiting the biocompatibility restriction, synthetic visual perception systems can be constructed to function as robotic eyes in artificial intelligence platforms such as autonomous vehicles.The use of a 3D construction, such as the one shown in Figure 1b, can eliminate the biocompatibility requirement from the innermost layers for such applications.These applications in turn can rely on flexible, high density, and nonorganic material-based ASDs to perform such complex computations.Low-dimensional materials are promising candidates for the implementation of such devices.One such material that immediately comes to mind in this context is graphene, [75] and indeed, it was used for demonstrating ASDs.In addition to its compatibility with complementary metal-oxide-semiconductor (CMOS) technologies and good thermal stability, [76] it has good potential for applications in flexible electronics due to excellent mechanical properties. [77]Moreover, as these materials are in general not classified as biocompatible, they would be relevant for the implementation of artificial retinas for robotic vision.ANNs that form a part of fully automated artificial systems are not subjected to the harsh restrictions dictated by compatibility with living organisms.
The first concept for utilizing graphene in ASDs is based on a vertical construction.In this construction, the 2D material is operated as an electrode or an interface rather than being an active memory layer.One example was demonstrated by Qian et al. [78] In their work, a large-area chemical vapor deposition (CVD) graphene in TiO x -based memristive devices was used to realize low switching power and nonlinear electric characteristics while maintaining good memory window, endurance, and retention properties.In addition, they fabricated flexible and transparent devices on poly(ethylene naphthalate) that shows excellent retention under mechanical bending, low switching current (1 μA), and high resistance.Another implementation by Lee et al. [79] took this concept a step further by using a graphene film with engineered nanopores integrated into a tantalum oxide-based lateral structure.The pores were used to allow controlled ion dynamics at the nanoscale, while the graphene film effectively blocked ionic transport and redox reactions.In this manner, the oxygen vacancies required for the resistive switching process [80,81] were allowed to pass only through the engineered nanosized pores, leading to effective modulation of the device performance.Another approach was based on the use of graphene as the active layer switching material still within a vertical structure.Park et al. [82] used this concept to demonstrate an artificial synapse constructed by crossing two yarns coated with reduced graphene oxide.Their artificial synapses could be operated without degradation during mechanical bending by implementing a 2 Â 2 cross-point array using coated yarns.Although far from being in the nanoscale regime, these devices demonstrate the possibility of integrating graphenebased artificial synapses for very flexible neuromorphic systems by electrochemical deposition over flexible fibers.
A different concept that designates graphene as the active material was shown over a lateral device structure.The main advantage in this approach is the possibility to unleash the full potential of a 2D material in 3D stacking of layers.Sharbati et al. [83] fabricated an electrochemical graphene synapse, where the electrical conductance is reversibly modulated by the concentration of Li ions (due to their high diffusion coefficient in bilayer graphene at room temperature) between the graphene layers via electrochemical intercalation.They demonstrated a low power synaptic device (<500 fJ per switching event) with good endurance and retention properties along with a linear and symmetric response.In addition, they demonstrated the occurrence of LTP and LTD in the 2D electrochemical synapse by the application of a series of intercalation and deintercalation pulses (50 pA, 10 ms) and could program over >250 distinct states.Furthermore, the conductance was stable under a 0.1 V reading bias (across the graphene channel) for more than 13 h with a 3.2% decrease in conductance at the end.The authors also performed an endurance test by cycling the device between two intermediate states for 500 cycles and observed small cycle-to-cycle variations.
Two-dimensional van der Waals materials, having a single or few layers, such as transition metal dichalcogenides and transition metal oxides, show intriguing electrical properties in homoor heterostructured configurations. [84]As such, another wellknown material is monolayer MoS 2 that was widely studied for 2D electronic, [85] optoelectronic, [86] spin-electronic, [87] thermoelectric, [88] and flexible applications. [89]Its incorporation into neuromorphic devices was thus quickly implemented.Several interesting works were published on memristive devices and transistors based on single-layer MoS 2 . [90]Another example demonstrated by Bessonov et al. [91] used vertically stacked MoO x / MoS 2 (x < 3) active layers with printed silver BEs and TEs on a plastic foil to demonstrate memristive and memcapacitive switches.Their devices had large and tunable electrical resistance ranging from 10 2 to 10 8 Ω with low programming voltages of 0.1-0.2V.A similar approach was demonstrated by Cheng et al. [92] to stack 1T phase of exfoliated MoS 2 nanosheets between silver electrodes.The authors demonstrated memristive behavior using this structure, which was attributed to electric fieldinduced lattice distortion in the 1T phase leading to resistant change.They further developed an ideal odd-symmetric memristor with odd-symmetric IÀV characteristics.
Another approach is based on the incorporation of a 2D material with proton-conducting electrolytes.Jiang et al. [93] used it to fabricate a synaptic device from poly(vinyl alcohol) (PVA) protonconducting electrolyte, laterally coupled, 2D MoS 2 electric double-layer transistor.Their device was structured in a lateral manner with two side gates and an additional modulator gate, all coupled to a source drain path via PVA.This configuration effectively formed a structure that functioned as an integrator, summing the combined effect of both driving and modulating inputs to spiking events.In this manner, the devices functioned as either AND orOR logic gates, based on the relative spike timing in different inputs.In addition, the authors used their device to demonstrate that when the spiking frequency was changed from 10 to 100 Hz, the emulated EPSC increased from about 100-500 nA.
Yang et al. [94] demonstrated a synaptic transistor based on quasi-2Dα-phase molybdenum oxide (α-MoO 3 ), which is a 2D transition metal oxide with good stability and compatibility properties.The structure relies on the ionic liquid 1-ethyl-3-methylimidazolium bis(trifluoromethanesulfonyl)imide (EMIM-TFSI) dropped on top of the device and functioned as the gate terminal as shown in Figure 10a.A schematic illustration of the device structure and measurement setup is shown in Figure 10b.The drain current (I D ) and gate leakage current (I G ) were measured as a function of the drain (V D ¼ 50 mV) and gate (V G ) potentials, respectively, at room temperature.These parameters in turn showed a large dependence on ambient atmosphere with different relative humidity levels (Figure 1c,d).In addition, the channel conductance could be set to different levels by application of consecutive voltage pulses with fixed voltage and width (2.5 V and 1 ms) for emulating synaptic potentiation (À1.8 V and 1 ms) and depression.The authors showed that after applying 200 pulses, the drain current was at the level of 3.5 nA and increased to about 4.7 nA after applying additional 50 positive pulses.The level then decreased back to 3.5 nA after the application of 50 negative bias pulses.
Within this framework, Sun et al. [95] demonstrated an interesting concept where Joule heating induced metal-insulator transition in a scalable MoS 2 synaptic device with very low energy consumption (%10 fJ).They further implemented a circuit with tunable excitatory and inhibitory devices to demonstrate sound localization via detecting an interaural time difference (i.e., the difference in arrival time of a sound between two ears; it is used to provide a cue to the direction of the sound source) by suppressing sound intensity-or frequency-dependent synaptic connectivity.The increased conductance in their devices (with a channel length of merely 50 nm and width of 200 nm) were attributed to a resistive heating effect on the semiconducting MoS 2 channel.They further implemented sound localization-like mechanism based on coincidence detection by measuring time difference and level difference using two distinct neural integration paths containing frequency selective inhibitory and excitatory synapses.
Another approach by Zhou et al. [96] used plastic substrates to fabricate indium-zinc-oxide flexible synaptic transistors with proton-conducting phosphorus-doped nanogranular SiO 2 electrolyte as the gate dielectric.The devices showed an EPSC, PPF, and long-term memory behaviors associated with synaptic activity.The laterally structured device used an ITO back gate to facilitate ion migration toward the channel/electrolyte interface when a voltage bias of %1.5 V was applied on the gate electrode.Their observed EPSCs were associated with the slow response of protons in the electrolyte film.
Alternatively, Wang et al. [97] implemented a device in which the active material was constructed out of a combination of both graphene and MoS 2 .In their work, layered stacking was used to effectively form Van der Waals heterostructures composed of graphene/MoS 2-x O x /graphene and fabricate robust memristors with good thermal stability.The devices exhibited good switching performance with an endurance of up to 10 7 cycles and a high operating temperature of up to 340 C.They used graphene (%8 nm thick) and MoS 2 (%40 nm thick) membranes that were mechanically exfoliated and deposited on SiO 2 /Si wafers.The main advantage in this construction is the realization of the atomically sharp interfaces between the switching layer and Reproduced with permission. [94]Copyright 2009, Wiley-VCH.
the electrodes.As the switching performance of memristors was shown to be largely affected by the roughness of the interface, [98] the heterostructured device showed a potential for high performance in comparison.The device showed repeatable bipolar switching, where the switching polarity was determined by the electroforming voltage polarity (i.e., positive/negative electroforming resulting in ON switching with positive/negative bias and OFF switching with negative/positive bias).
The use of low-dimensional materials is by no means limited by the 2D class.Several works used 1D carbon nanotubes in the construction of ASDs.One of the main features that make nanotubes so attractive is the very high mobility (%269 cm 2 V À1 s -1 ).Feng et al. [99] constructed neuromorphic devices based on printed single-walled carbon nanotube (SWCNT) transistors.Their device had a saturation current of À442 μA at V DS ¼ À1.2 V and V GS ¼ À0.8 V with linear properties.In addition, it showed high mobility, high ON/OFF ratio, and low leakage current at low operation voltage (AE1 V).The nanotubes were entirely embedded into the gate dielectric layer for the following two reasons: 1) hole concentration of the SWCNT channel can be effectively tuned in a very low voltage range and 2) utilizing a very low leakage gate-all-around structure.The authors used their device to demonstrate EPSCs found in biological neural systems.The SWCNT transistors were initially set to HRS by using a base presynaptic voltage of 0.6 V applied on the presynaptic input.Two hundred presynaptic voltage pulses of 24 ms width were then applied to the presynaptic input at different voltages.First, the EPSC amplitude increased nearly linearly from below 2 nA to approximately 0.1 μA after approximately 30 pulses for voltages of À0.2 and À0.5 V, until reaching a maximum of about 0.13 and 0.20 μA, respectively.
Kim et al. [100] took this concept a step further and showed pattern recognition with carbon nanotube synaptic transistors by using an adjustable weight update protocol.Although this work aimed to demonstrate an application rather than a whole new device concept, it is worth mentioning in this context.They fabricated a synaptic transistor based on highly purified, 99% semiconducting carbon nanotubes, which can provide adjustable weight update linearity and variation margin.A synaptic function was emulated by the configuration of three carbon nanotube (CNT) transistors that formed an inverter driving a synaptic transistor.
In their construction of a neural network, each output neuron was connected with 28 Â 28 synaptic transistors in parallel, where each synaptic transistor generates the postsynaptic current based on the channel conductance.The sum of the postsynaptic currents is accumulated by an integrator, and finally, the output neuron generated a postsynaptic spike through a waveform generator depending on a predefined threshold.To demonstrate pattern recognition using their network, the authors used the modified National Institute of Standards and Technology (MNIST) database, which consists of handwritten numbers that are 28 Â 28 pixels in size.The network training phase consisted of inputting the full MNIST training database, which consists of 60 000 digits, into the system with each input neuron connected with one pixel of the image.In this manner, the input neurons emit presynaptic spikes with timings that are proportional to the image pixel intensity.These input presynaptic spikes generate postsynaptic currents based on the synaptic weight of each synaptic transistor, which in turn are integrated by the output neurons.Finally, the output neuron, with the highest postsynaptic current, fires a postsynaptic spike.After completing the learning process, the network was successfully tested on the MNIST test database, which consists of 10 000 images, and it showed a recognition rate of about 60% after 105 learning cycles.
Yet another extension of this concept was shown by Esqueda et al. [101] They demonstrated synaptic transistors constructed of aligned CNT for large-scale neuromorphic applications.They utilized charge trapping in the high-k dielectric layer of topgated CNT field-effect transistors (FETs) to enable gradual analog programmability of the channel conductance with a large dynamic range (i.e., on/off ratio).The high uniformity of the aligned CNT allowed constructing devices with reliable synaptic behavior and robustness when operated through pulsing schemes.The issue of self-alignment and semiconducting purity is critical for optimizing device performance and reliability in CNT-FETs. [102]The authors therefore used an evaporation-driven process, named floating evaporative self-assembly, to fabricate their devices. [103,104]The drain current of a P-type CNT FET with a channel width to length ratio (W/L) of %60 μm/1 μm for a drain bias of À1.0 V showed a distinct counterclockwise gate hysteresis.This was attributed to a dynamic screening of the electric field due to charge injection/emission (i.e., trapping/detrapping) at the interface between the CNTs and the 4.6-nm-thick HfO 2 gate dielectric.The synaptic properties (potentiation and depression) of the CNT FETs were demonstrated using a series of gateto-source (V gs ) and drain-to-source (V ds ) voltage pulses and yielded roughly a two order of magnitude change in channel conductance.In addition, they also tested long-term retention of synaptic weights through the time-dependent sampling of the drain current after programming of various states.Their results showed only a small loss of the extreme states that correspond to the largest/smallest programmed channel conductance occurs after about 10 3 s.
The incorporation of zero-dimension materials (quantum dots [QDs]) was also demonstrated as beneficial to improve the properties of ASDs.Yan et al. [105] implemented memristors with progressive conduction tuning based on a vertical structure while embedding graphene oxide QDs into a Zr 0.5 Hf 0.5 O 2 matrix.The conductance of the device was adjusted through a series of pulses with an amplitude as low as 0.6 V and a width of 30 ns.They demonstrated that even low energy pulses can realize a linear conductance regulation with good accuracy.The authors further demonstrated that conduction modulation could be performed using two main approaches.One is current gradual change, and another is current abrupt change.Under a wider pulse width or higher amplitude, abrupt resistive switching was always observed.On the other hand, voltage pulses with smaller amplitude and narrower width change the current slowly, and as a result, offer more intermediate resistive states for multilevel memory functionality.In fact, the concept of embedded nanostructured metals in the switching matrix of memristors was extensively studied mainly in an attempt to increase electrical parameter uniformity and multilevel operation. [106]These memristors help to reduce operating voltages and provide better retention and higher on/off ratios due to localization of electric fields.Some examples are based on Ag, [107] Au, [108] and Hf. [109]

Concluding Remarks
A realistic implementation of a bionic retina must comply with some key restrictions mandated by an in vivo operating module.Artificial devices comprising each part of this system must be evaluated first and foremost in this light.Nevertheless, such restrictions may be completely lifted when targeting robotic vision systems.The first aspect one must consider would therefore be energy consumption and thermal dissipation.Most of the biocompatible devices discussed herein are large in size and operate using large currents when compared to inorganic material-based ASDs.Low-dimensional devices, on the other hand, clearly have an advantage in this aspect due to their small dimensions and low power consumption.In addition, the active materials used to construct them are CMOS compatible.The construction of robotic vision platforms, capable of large-scale image processing operations, should be based on such devices.
However, the electrical properties of current biocompatible devices do not necessarily render them completely irrelevant, considering that the degree of functionality expected from a bionic system is much lower than that of a robotic system (i.e., much simpler and rudimentary image processing).In addition, connecting the output stage directly to the optical nerve would require a surgical procedure and the total number of connections thus cannot be very large.As a result, the total number of ASDs needed for each hidden layer would be relatively low and so will the total size and power resources consumed by each layer.Basic ANNs, operating with few tens of devices, were already demonstrated as being able to perform simple image classification tasks, which may be sufficient for the operation of a prosthetic eye.
The second aspect would be the compatibility of different materials used in the entire system.For the purpose of highdensity 3D integration, a single process would be preferred.The compatibility aspect should also be considered in the light of reliability of the entire integrated system.Any conflicting issues with materials used for each layer should be avoided to produce a fault-tolerant system with statistically reproducible results.The choice of devices for each layer should be done accordingly in addition to the required functionality by each layer.An overall comparison of the ASDs discussed in this report according to materials, structure, and electrical properties is shown in Table 1.
The third aspect to consider is related to the dynamic properties of the entire system.As it is intended to operate virtually instantly (produce results within a single cycle), the first element Table 1.A comparison of ASDs with respect to materials used, structure, attributes, and operating conditions.to determine it would be the switching speed of the photoresponsive layer.The restive switching process itself was shown to occur within nanosecond speeds; [110] the bottleneck therefore may be in the resetting process of the array between image acquisitions.In this light, the use of either a volatile or nonvolatile photo-responsive material come to play.The main advantage of a volatile device would be bypassing the need to implement a reset mechanism.In this manner, images are continuously captured and updated by the photo-responsive layer, much like in the biological system, and it should be used in a bionic retina.A robotic vision system, on the other hand, would benefit more from being based on nonvolatile devices because the incorporation of a reset mechanism would be straightforward.In addition, the nonvolatility property offers higher fault tolerance to noise and power surges.
Figure 1b,d,e depicts a 3D placement of crossbar arrays (highlighted in blue) intended to implement rudimentary image processing computations.Put together, they form a series of Dan Berco received his bachelor's and master's degrees from the Technion Institute of Technology, Israel, and his doctoral degree from Electrical Engineering Department of the National Chiao-Tung University, Hsinchu, Taiwan.He is currently a research fellow in the Electrical and Electronic Engineering Department of Nanyang Technological University, Singapore.His research interests include computational plasmonics, nanoionic-device physics, and neuromorphic computing.Diing Shenp Ang obtained both his B.Eng. (Hons.) and Ph.D. degrees in electrical engineering from the National University of Singapore.He is currently an associate professor in the School of Electrical and Electronic Engineering, Nanyang Technological University, Singapore.His research interests include transistor reliability physics and characterization.Also he is interested in the applications of nano-characterization techniques, nanostructures, and resistance switching devices.

Figure 2 .
Figure 2. a) Schematic illustration of a retina.b) A 3D structure of the device along with their chemical structure.c) Functional principles of STP (left) electrochemical doping-induced LTP (middle) and ferroelectric switching-induced LTP.d) Signals in the array recorded after NIR exposure for 20 s. e) Remnant signals after 600 s. f) Remnant signals after 1800 s; exposed pixels are indistinguishable from unexposed pixels.g) Signals in the array recorded after green light exposure for 4 s.h) Remnant signals after 600 s. i) Remnant signals after 1800 s, indicating that the array retained at least 65% of the initial level.Reproduced with permission.[37]Copyright 2018, Wiley-VCH.

Figure 3 .
Figure 3. a) Schematic of the memristor measured under light illumination.b) IÀV characteristics of a device measured at different illumination intensities (from 0 to 1.29 μW cm À2 ).c) V 1 (voltage required to switch the device to the same LRS) and V 2 (voltage required to hold the device at the LRS) dependence on illumination intensity.d) Illustration showing that light illumination can inhibit iodine ion and vacancy formation under bias.e) Potentiation curves at different illumination conditions (from 0 to 0.38 μW cm À2 ) with electrical pulses (0.5 V, 1 ms, 0.1 ms interval).f) Final conductance after 200 electrical pulses, at different illumination intensities.g) Illustration showing that light illumination can accelerate iodine ion and vacancy annihilation without bias.h) Conductance retention curves at different illumination intensities for the same initial conductance value.i) Characteristic decay time constant of the device as a function of illumination intensity.j) Conductance retention curve of a device during which light illumination (1.29 μW cm À2) was alternately applied.k) Evolution of the conductance when stimulated with electrical pulses (1 V, 10 ms, 1 ms interval), in the dark and under light illumination (1.29 μW cm À2 ).Reproduced with permission.[38]Copyright 2018, American Chemical Society.

Figure 4 .
Figure 4. a) Self-rectifying I-V curves showing a photodiode behavior.b) Photoconductivity in a W/MoS 2 /p-Si device with UV irradiation.c) I-V characteristics after 15 sweeps.d) HRS and LRS reads at 0.1 V indicate a complete decay of a volatile memory state after %150 s. e) Photonic potentiation and electric habituation.Reproduced with permission.[39]Copyright 2018, Wiley-VCH.

Figure 5 .
Figure 5. a) Schematic depiction of black phosphorus memory device with a 30 nm Al 2 O 3 charge-trapping layer.b) Atomic force microscope image showing a thickness of about 10 nm.c) Ambipolar conduction as a function of back-gate bias, indicating a 0.3 eV bandgap.d) Typical hysteresis window of the device showing a leakage current below 20 pA.e) Illustration of a modulation using a broadband mid-IR transmitted through a black phosphorus flake.f)Fourier-transform infrared spectroscopy (FTIR) transmission extinction versus photon energy normalized to zero bias.a-d) Reproduced with permission.[48]Copyright 2016, American Chemical Society.e,f) Reproduced with permission.[49]Copyright 2016, American Chemical Society.

Figure 6 .
Figure 6.a) Schematic illustration of a conductive path comprising a cluster of oxygen vacancy defects formed after soft electrical breakdown.b) During illumination, photons excite these interstitial oxygen ions and the latter migrate toward the conductive path (marked by a white arrow).c) I-V curves for the SBD, post-SBD, and post-white-light illumination.d) Temporal current behavior showing a decrease in the current upon illumination by white light.e)Speed of the negative photoconductivity response of the ZrO x film at two different white light intensities.f) Speed of the negative photoconductivity response of the oxygen-deficient and oxygen-rich HfO x films.Reproduced with permission.[53]Copyright 2018, AIP Publishing.

Figure 7 .
Figure 7. a) Schematic illustration of Mg/Ag-doped chitosan/Mg device structures.b) Flexible and biocompatible devices.c) I-V characteristics of the Mg/Ag-doped chitosan/Mg resistive switching memory device.d) Data retention characteristics of LRS and HRS states under continuous readout voltage at room temperature.e) Switching endurance of the device measured for up to 60 cycles.f) Cycle-to-cycle statistical cumulative distribution of the SET and RESET voltages.Reproduced with permission.[57]Copyright 2015, Wiley-VCH.

Figure 8 .
Figure 8. a) Nonlinear responses of lignin-based synaptic device, IÀV characteristics of the device during five consecutive positive and negative sweeps.b) Conductance behavior during consecutive sweeps.c) Temporal current response as a function of applied bias pulses.d) Current variation after 50 consecutive negative pulses (À0.7 V, 100 ms), followed by 50 consecutive positive pulses (þ0.7 V, 100 ms).e) I-V curves for three different devices on silk substrate: W/MgO/Mo, W/ZnO/Mo, and W/MgO/ZnO/Mo.f) 300 circles of DC sweeping I-V curves of the W/MgO/ZnO/Mo memristor.g) Endurance characteristics of the W/MgO/ZnO/Mo memristor.h) Retention characteristic of the W/MgO/ZnO/Mo memristor under 0.1 V read voltage at room temperature.i) I-V characteristics of Mg/collagen/ITO/PET in flat state.j) I-V characteristics of Mg/collagen/ITO/PET under tensile bending with a radius of curvature of 7 mm.a-d) Reproduced with permission.[60]Copyright 2017, American Chemical Society.e-h) Reproduced with permission.[61]Copyright 2018, Royal Society of Chemistry.i,j) Reproduced with permission.[63]Copyright 2018, Wiley-VCH.

Figure 9 .
Figure 9. a) Schematic depiction of a 10-μm OEIP with source (S), waste (W), and target (T) electrodes on electronically conducting PEDOT:PSS connected by the cation-selective channel (pink).ACh (red) is delivered at the 10-μm channel outlet, where it spreads by diffusion to the SH-SY5Y cells (green) cultured on top of the target electrode.b) Temporal dynamics of Ca 2þ signaling in cells located at 50 μm and 150 μm from the 10-μm channel outlet on delivery of ACh.Voltage pulses of 20 V for indicated times generated Ca 2þ oscillations.c) Schematic illustration of a pnp-IBJT with an emitter, collector, and base channels and a junction consisting of a neutral polymer gel electrolyte.The conductive PEDOT:PSS electrodes are covered by electrolytes and injected and/or extracted ions from the terminals.In the active mode (V EC ¼ 10 and V EB ¼ 4 V), the base supplies the junction with Cl À .Increased conductivity results in ACh transport from emitter to collector.In the off-mode (V EC ¼ 10 V and V EB ¼ À1 V), the base depletes the junction of Cl À , and ACh delivery stops due to decreased conductivity.d) Intracellular Ca 2þ recording of ACh-stimulated SH-SY5Y cells cultured on the collector terminal.Turning the base on/off regulates ACh delivery.a,b) Reproduced with permission.[66]Copyright 2009, Wiley-VCH.c) Reproduced with permission.[68]Copyright 2011, American Chemical Society.d) Reproduced with permission.[67]Copyright 2010, National Academy of Sciences.

Figure 10 .
Figure 10.a) Optical image of the quasi-2D α-MoO 3 three-terminal synaptic device.b) Schematic illustration of the device structure and measurement setup.A small DC voltage (V D ¼ 50 mV) was applied between the source and drain electrodes; gate voltage (V G ) applied to the gate electrode and corresponding drain (I D ) and gate (I G ) currents monitored.c) I D dependence on the gate voltage.d) I G dependence on the gate voltage respectively, under different relative humidity conditions (vacuum, 18.6%, 36.2%, and 45.1%).e) Schematic depiction of the transistor structure corresponding to a positive gate voltage.The applied electric field drives the protons and hydroxyls, which dissociate from H 2 O adsorbed in the ionic liquid, in the opposite direction.The protons are adsorbed at the α-MoO 3 channel surface and then injected into the α-MoO 3 lattice, resulting in an increase in the channel conductance.f) Schematic of the transistor structure corresponding to the negative gate voltage.The protons are extracted and desorbed from the α-MoO 3 channel surface with the accumulation of hydroxyls, resulting in a decrease in the channel conductance back to the initial state.Reproduced with permission.[94]Copyright 2009, Wiley-VCH.