In this talk, I will share our work on the ionic electrochemical synapses, whose electronic conductivity we control deterministically by electrochemical insertion/extraction of dopant ions into/out of the channel layer. This work is motivated by the need to enable significant reductions in the energy consumption of computing, and is inspired by the ionic processes in the brain. Proton as the working ion in our research presents with very low energy consumption, on par with biological synapses in the brain. Our modeling results indicate the desirable material properties, such as ion conductivity and interface charge transfer kinetics, that we must achieve for fast (ns), low energy (< fJ) and low voltage (<1V) performance of these devices. Importantly, the conductance change in these electrochemical devices depends non-linearly on the gate voltage, due to field-enhanced ion migration in the electrolyte, and charge transfer kinetics at the electrolyte-channel interface. We are leveraging these intrinsic nonlinearities to emulate bio-realistic learning rules deduced from neuroscience studies, such as spike timing dependence of plasticity and Hebbian learning rules. Our findings provide pathways towards brain-inspired hardware that has high yield and consistency and uses significantly lesser energy as compared to current computing architectures.

This talk describes state-of-the-art organic neuromorphic devices and provides an overview of the current challenges in the field and attempts to address them. I demonstrate two device concepts based on novel organic mixed-ionic electronic materials and show how we can use these devices in trainable biosensors and smart autonomous robotics.

Next to that, the process of neural network training can be slow and energy-expensive due to the transfer of weight data between digital memory and processor chips. Neuromorphic systems can accelerate neural networks by performing multiply-accumulate operations in parallel using non-volatile analogue memory. However, the backpropagation training algorithm in multi-layer (deep) neural networks requires information - and thus storage - on the partial derivatives of the weight values, preventing easy implementation in hardware.

In this talk I will highlight a novel hardware implementation of the well-established backpropagation algorithm that progressively updates each layer using in situ stochastic gradient descent, thus avoiding this storage requirement. We experimentally demonstrate the in situ error calculation and the proposed progressive backpropagation method using a multi-layer hardware implemented neural network based on organic EC-RAM, and confirm identical learning characteristics and classification performance compared to conventional backpropagation in software.

Biointegrated neuromorphic hardware holds promise for new protocols to record and regulate signaling in biological systems. Traditional neuromorphic systems based on silicon have limited bio-integration potential due to their high circuit complexity, poor biocompatibility, and low energy efficiency. Organic mixed ionic-electronic conductors (OMIECs) offer a potential solution to these limitations. Because of their structural kinship with biomolecules and coupled ionic-electronic transport functionalities, OMIECs are an excellent choice for bridging electronics and biology, enabling energy-efficient signal transduction. In this presentation, I will explore the use of OMIECs to develop organic electrochemical neurons with ion-modulated spiking capabilities. We will discuss the ease of their integration with biological neurons and demonstrate their operation as an event-based sensor, transducing biochemical signals to electrical signals that can be used to actuate/stimulate biological nerves. These soft, flexible organic electrochemical neurons and synapses operate at low energy and respond to multiple stimuli, signaling a new era for closed-loop regulation of physiology.

The growing demand for data-driven applications in every layer of society results in traditional von-Neumann architecture with disjoint memory and processing to be at its limit, as it suffers from huge memory latency and limited data bandwidth, which is further intensified by the scaling limits of silicon transistors.
A new roadmap for beyond CMOS is under discussion and this is driving interest in utilizing unique characteristics of emerging devices for information processing and memory[1]. One of the challenges is to identify emerging devices that can implement computing functions and architectures more efficiently than CMOS and Boolean logic.
Neuro-inspired chips that mimic the behavior and the efficiency of biological networks of neurons and synapses are receiving particular attention, as most of the AI related problems rely on neural network mathematical models. Reducing data movement and parallelizing vector matrix multiplications are key aspects in establishing more energy-efficient AI hardware. Device technologies based on metal-oxide HfO2 and VO2, as well as phase change memories (PCM) based on GeSbTe and Sb2Te3 are perfect candidates for the realization of synapse and neuronal functions in these artificial neural networks.
In this talk I will give an overview on our approach to accelerate deep neural networks for inference and training applications using Resistive Random-Access Memories (ReRAM)[2] and PCM[3], as well as on the use of oscillating neural networks based on VO2 oscillators for solving optimization problems[4].
 

Neuromorphic sensing and computing exploits signal encoding and computing inspired by biological brains, down to spiking signal representation. Brain efficiency is overwhelming. the human brain consumes about 20W of power but is capable of remarkably sophisticated cognitive tasks with very low latencies. The visual system, for example, can explore highly complex scenes and be able to detect specific objects, partially occluded, and quite quickly. In this presentation, we will show event cameras that mimic biological retinae. They do not generate frame sequences but provide a flow of continuous spikes (which we call events), each with a pixel coordinate, that represent the dynamic changes in a visual scene. These address events become available with sub-microsecond latency and can be processed by a spiking neural neural immediately. This way, a compound system with an event camera and a hardware spiking neural network can perform object recognition with sub-millisecond latency. For the spiking processing, we will show chips that exploit nanoscale memristor devices as synaptic devices.

This talk presents the NimbleAI approach, which leverages the key principles of energy-efficient light sensing in eyes and visual information processing in brains. Inspired by these principles, NimbleAI is creating an integral sensing-processing neuromorphic chip that builds upon the latest advances in 3D stacked silicon integration. In NimbleAI, a frugal always-on sensing stage builds basic understanding of the visual scene and drives a multi-tiered collection of highly specialised event-based pre- and post-processing kernels and neural networks to perform visual stimuli inference using the bare minimum amount of energy. Since manufacturing a full 3D testchip is prohibitively expensive, NimbleAI prototypes key components via small-scale 2D stand-alone testchips. This cost-effective use of silicon allows us to readjust direction during project execution to produce silicon-proven IP and high confidence research conclusions. To enable the global research community to use our results, we will deliver a prototype of the 3D integrated sensing-processing NimbleAI architecture with the corresponding programming tools. The prototype will be flexible to accommodate user IP and will combine commercial neuromorphic chips and NimbleAI testchips.

In recent years, the insulator-to-metal transition in Vanadium Dioxide (VO₂) micro-resistors has been harnessed to design spiking, sensitive neurons [1-6], implemented through compact circuits with minimal components number. VO₂ micro-resistors feature "S-shaped" I-V characteristics with a negative-differential resistance (NDR). Using a current source, the current flowing through the micro-resistor can be set within this region, sot that the device enters an astable regime and quickly oscillates between its metallic and insulating phases, producing voltage spikes.

These sensitive neurons can encode various stimuli including temperature, pressure, RF signals, or visible light into the spikes rate. The transduction is realized either by (1) a dedicated sensor whose current output is fed to the VO₂ device [1-2] or (2) the modulation of the VO₂ device electrical characteristics by the stimulus [3-6], both options affecting the spike rate. However, so far, all proposed sensory neurons – including our previous work [6] – function in an "always spiking" fashion, producing continuous spike trains where information is solely encoded into the time period of consecutive spikes. These sensors are meant to be interfaced with Spiking Neural Networks (SNNs): when using such encoding, the sensing/processing system does not benefit from the full energy saving potential SNNs can offer when activated in a sparse fashion [7]. In addition, plain rate-coding is only one of the many neural codes observed in biological neurons [8], meaning present implementations of VO₂ sensitive neurons do not reproduce this complexity.

In this work, we propose a temperature-sensitive VO₂ neuron operating just below the limit of its spiking regime, in which temperature is encoded into several characteristics of spike bursts rather than continuous spike trains. This bursting phenomenon could be attributed to a combination of noise sources, both external and intrinsic to the VO₂ device. In our case, the current is set right under the NDR region of the device, using a NMOS transistor in saturation regime (making the sensory neuron a 1T-1R system). The VO₂ micro-resistor will then sporadically enter spiking regime, through stochastic resonance with external or internal noise (e.g. on the transistor current, supply voltage, or thermal noise on the resistor), triggering the start of a spike burst. After a few spikes, the burst terminates: this could be due to intrinsic cycle-to-cycle variations in the VO₂ electrical characteristics (modeled and observed in previous work [6][9][10]).

We study the evolution of several bursts characteristics over the 42 – 45°C range: bursts duration, inter-burst intervals, spike count per burst, frequency of burst occurrence, inter-spike interval (ISI) in burst, and mean firing rate. For the four last parameters, we find trends similar to those observed in biological thermoreceptors of insects. Indeed, such receptors switch from regular to bursty spiking in certain noxious temperature ranges; their burst frequency, spike count per burst and mean firing rate increase with temperature, while their ISI decreases [11][12]. In VO₂ neurons, this can be explained by the modulation of the I-V characteristics with temperature: the S-shaped region moves closer to the fixed current, making the device more likely to enter spiking regime upon current fluctuations.

This bursting behavior takes place in a narrow temperature range: above a certain threshold temperature, the VO₂ neuron enters back into the always-spiking regime, in which temperature is solely encoded into the period of continuous spikes. We measure 10 different VO₂ devices and find that this threshold temperature varied from up to several degrees from neuron to neuron, a disparity which is also observed among biological neurons in single specimens [11].

Our findings pave the way for a new type of more biologically plausible, energy efficient sparse encoding of stimuli into phase-transition neurons.

Neurodynamic behavior of artificial neuron circuits made of Mott memristors [1] provide versatile opportunities to utilize them for artificial sensing [2,3]. In these realizations, usually a sensor is connected to an artificial neuron or oscillator circuit to generate a spiking output encoding the stimulus, which is later carried to a spiking neural network for further processing. By directly conducting such signals encoding stimuli into the nervous system, an artificial sensory input can be created. We explore possibilities to realize an auditory sensing circuit aiming future applications in cochlear implants, motivated by small size and energy-efficient spike generation capabilities of VO₂ oscillator circuits. The sensory part of the circuit can be realized with micro-electromechanical systems (MEMS), also applicable as bio-inspired acoustic sensors [4,5].

In this work, a MEMS cantilever is connected to a VO₂ nanogap Mott memristor [6] -based oscillator circuit, capable of neural spike emission. The MEMS cantilever can realize frequency selective sensing of sound waves due to its sharp resonance. Electrical signal of the cantilever serves as an input for the VO₂ oscillator circuit after proper conditioning. The cantilever is excited with mechanical stimuli in the range of ~10 nm, which is realistic in the human ear [7]. As a result, the oscillator emits a train of spikes with a well-defined frequency, which can be tuned to the desired frequency domain of typical spiking rates in the nervous system [8], via the proper selection of passive elements in the oscillator. By the addition of further passive elements to the oscillator, the spiking waveform can be brought to a bipolar form. This is a beneficial property, since electric signals carried into the nervous system should have bipolar characteristics to exclude the possibility of charge accumulation near the end of the implanted electrodes [9]. Due to the internal dynamics of the oscillator part, the frequency of oscillation encodes the amplitude of the stimulus – similarly to processes of natural hearing [10]. The proposed circuit serves as a proof-of-concept demonstration of a Mott memristor-based auditory sensing unit for future cochlear implants.

Novel nanoelectronics technologies have been utilized to develop systems capable of running various artificial neural network (ANN) architectures, presenting the potential to power future edge devices. Memristor-based ANN architectures may be particularly appealing due to their in-memory processing capabilities, high throughput, and energy efficiency. We hypothesize that systems capable of directly processing complex-valued signals (e.g., those originating from the external world), could be more efficient than systems that necessitate the separation and independent processing of amplitude and phase information, which is the case form most conventional systems. Here, we introduce innovative nanodevices utilizing W/Ba-doped HfO₂/Nb:SrTiO₃ stacks, which exhibit simultaneous conductance and capacitance changes during both set and reset processes, establishing them as both memristive and memcapacitive devices. These interface-type switching devices leverage multiple impedance states and allow us to gradually alter set and reset kinetics. This concept could provide a foundation for the development of systems that directly implement complex-valued neural networks (CVNNs). This allows us to reduce the circuit complexity, number of training epochs, and accordingly the overall energy efficiency to distinguish them from any traditional two-dimensional real-valued neural networks (RVNN). Overall, this study paves the way for the development of energy-efficient neuromorphic hardware.

Memristive devices hold promise to improve the scale and efficiency of machine learning and neuromorphic hardware, thanks to their compact size, low power consumption, and the ability to perform matrix multiplications in constant time. However, on-chip training with memristor arrays still faces challenges, including device-to-device and cycle-to-cycle variations, switching non-linearity, and especially SET and RESET asymmetry [1], [2].

To combat device non-linearity and asymmetry, we propose to program memristors by harnessing neural networks that map desired conductance updates to the required pulse times. With our method, approximately 95% of devices can be programmed within a relative percentage difference of ±50% from the target conductance after just one attempt. Moreover, our neural pulse predictor demonstrates a significant reduction in memristor programming delay compared to traditional write-and-verify methods, particularly advantageous in applications such as on-chip training and fine-tuning.

Upon deployment, the neural pulse predictor can be integrated into memristor accelerators, predicting pulses with an O(1) time complexity while utilizing a minimal fraction of the available memristor arrays, reducing hardware overhead compared with previous works [3]-[6]. Additionally, multiple networks can be trained to operate in parallel and enhance precision across various conductance ranges.

Our work contributes significantly to the practical application of memristors, particularly in reducing delays in memristor programming. This work also offers a fresh perspective on the symbiotic relationship between memristors and neural networks and sets the stage for innovation in memristor optimizations.

The emulation of brain functionalities through solid state electronic systems requires billions of dynamical components reproducing the time varying properties of neurons, synapses, dendrites, etc. (as exemplified in TOC figure on the left). Memristor devices that rely their electrical operation on ion movement, interestingly akin to the chemistry of neural processes, hold the potential to make it possible thanks to their nanoscale dimension and short- and long- term memory of past electrical signals. Indeed, voltage pulses modify the inner layer of their metal/oxide/metal structure through formation of ionic conductive filaments (CFs) which short the electrodes and modify the device overall conductance. The fine understanding of conductance dynamics of memristive devices is therefore crucial for the final goal of an electronic brain.

In this framework, we characterized prototypical Pt/SiO_x/Ag devices belonging to the class of electrochemical metallization memristors (the top right part of the TOC figure sketches the device structure and switching mechanism). Upon voltage application, Ag⁺ are injected from the electrode, migrate and deposit into a growing Ag CF that eventually shorts the two electrodes. As a result of these sequence of processes, the device reacts to a voltage pulse (grey area of the bottom right panel of the top figure) with an abrupt current upward jump (blue plot in the bottom right part of the TOC figure). The current jump is delayed by the moment of the voltage application by a certain time (called delay time), which is characteristic of the filament growth process. As soon as the voltage is released, the current gradually decreases, within a relaxation time, until a final drop to the pristine value occurs because of the atomic surface diffusion that disconnects the CF.

In the present work, we show that these timescales are at the base of the use of electrochemical devices for the emulation of neural functions. In particular, we identify prototypical dynamical features which can be useful for emulation of brain functions, as stimulated in response to sequence of pulses. We analyze such features as a function of the pulse sequence parameters (voltage, pulse width and inter-pulse intervals) so to draw a comprehensive picture of the device programming possibilities. We also describe the results with reference to the basic physical mechanism responsible for the CF growth and self-dissolution.

Memristor-based multi-level memory systems are gradually breaking through the limitations of traditional binary logic systems in terms of critical figures of merit, such as low power operation, operating speed, and circuit complexity. Logic gates and functions have been demonstrated in various ways, confirming their potential as a useful emerging computing device. However, stateful logic's spatio-temporal efficiency is not good enough to supplant traditional computing technologies due to the random formation of conducting filament (CF). We present inter-quantized conductance (QC)-state switching in a highly reproducible and stable quaternary QC-state memristor based on Al/iPrP-PDI/ITO. Memristor shows excellent bipolar switching performance of high continuous cycling with large ON/OFF ratios and long retention time with low operating voltages, as well as being remarkably reproducible over two years. The switching parameters' temporal and spatial variability was studied using 880 cycles across ten devices and found to be highly reproducible [1]. Furthermore, during SET/RESET cycles, discrete quantized conductance states were detected in I-V traces due to the development of atomic point contacts in the conducting filaments. Compliance current (I_c) and stop voltage (V_s) were used to control the CF. QC states were analyzed statistically, and three distinguished QC states, i.e., 1G0± 0.5G0, 3.5G0± 0.5G0, 7.5G0± 0.5G0, were found to be reproducibly available as stable memory levels. Controllable interstate switching among these QC states was also observed. Control on various switching configurations, such as switching ON to different QC-states, switching OFF from QC, and controlled inter-state switching from one QC-state to another, has been demonstrated by imposing current compliance and stop-voltage, showing potential for implementation in multi-level logic. Overall, iPrP-PDI has the potential to be an excellent choice for durable, high-performance, and highly dense multi-level memory devices for practical applications.

Keywords: Quantized conductance; organic semiconductors; isopropyl phenyl-Perylenediimide(iPrP-PDI); multi-level memory; inter-state switching.

With the advent of neural network tools like ChatGPT and DALL·E the energy demand for neural network servers continually rises to facilitate the neural network needs of people across the globe. Neuromorphic devices can drastically lower the energy demand of such neural network algorithms. Organic neuromorphic devices based on the bio-compatible conductive polymer PEDOT:PSS provide the additional capability to interface with living tissue and microfluidic systems. However, organic neuromorphic devices suffer from self-discharge caused by parasitic electrochemical reactions. These redox shuttle reactions are modeled by combining two-phase charge transport models with electrochemical self-discharge models. The successful implementation of these models allows for the accurate and extensive simulation of single-device discharge behavior and its influencing factors.[1]

Implementation of the single-device simulation into a simulated single-layer network reveals a 100 % accurate prediction over ten hours, even under significant weight drift. For multi-layer networks, however, the prediction performance degrades significantly after 20 minutes due to the weight deterioration caused by the self-discharge. Periodic reminder pulses are necessary to retain network performance. [2]

The necessity for complex compensation mechanisms of self-discharge can be eliminated in spiking neural networks (SNN). Inspired by biology, spiking neural networks implement local and always-on learning. When built with organic neuromorphic devices, these networks constantly relearn and reinforce forgotten states. Implementing an accurate surrogate model for the organic neuromorphic devices into a Brian 2 simulation enables the performance analysis of a spiking neural network on organic neuromorphic hardware. The surrogate model was derived from the high-resolution charge transport model of the previous simulations. 28 x 28 MNIST images were evaluated with a two-layer network. The network achieved competitive recognition results for its size. Networks using devices with self-discharge even achieve higher prediction accuracy than ideal devices without self-discharge. Online learning with idle rates of up to 90% can keep the network performance steady and close to the initial accuracy.[3]

These results reinforce the potential of organic neuromorphic devices for use in brain-inspired computing. Possible avenues for device integration include targeted drug delivery, bio-interfacing and sensing applications, and close integration with multi-electrode arrays.

Metal-Insulator-Metal (MIM) structures can exhibit memristive properties under appropriate biasing conditions. This phenomenon is based on the resistive switching effect, which is characterized by a pinched hysteresis I-V loop associated with reversible changes in the device resistance caused by an electrical stimulus. This technology is currently under extensive investigation for a wide number of applications such as non-volatile memories, digital logic circuits, hardware security, and brain-inspired computing architectures. In recent years, the influence of electrical stress on the reliability of memristors has been investigated [1,2]. However, the existing literature is limited to the analysis of physical damage resulting from electrical stress, especially in connection with catastrophic dielectric breakdown in resistive switching devices [3,4]. But several questions remain unclear, specifically about the impact of electrical stress and polarity effects in filamentary MIM-based memristors.

This work investigates the consequences of catastrophic damage in TiN/Ti/HfO₂/W memristors subjected to electrical stress. The analysis includes physical inspections and compositional evaluation of the device, employing scanning electron microscopy (SEM) (Fig.1) and energy-dispersive X-ray spectroscopy. The impact of voltage polarity on the observed effects is also discussed. Results indicate that under high negative voltages applied to the TiN top electrode, a worm-like pattern emerges over the active area of the devices (Fig. 1(a)). Additionally, occasionally, more severe damage occurs, characterized by disconnection and melting (Fig. 1(c)). In contrast, no damage is observed in devices subjected to positive polarity stress. In summary, this study contributes to the understanding of the consequences of catastrophic dielectric breakdown in TiN/Ti/HfO₂/W memristors under electrical stress, shedding light on the dynamic interplay between voltage polarity and the physical integrity of the devices.

Fig. 1. SEM images of the damage typologies observed during the physical inspection of the devices subjected to electrical stress under a DC voltage sweep with negative polarity. (a) The most common damage pattern observed (“the worm”). (b) Damage observed at the edge of the top metal line of the device. (c) Image of a melted device, leading to an electrical disconnection.

Valence change memories, where the change in resistance is based on the oxygen ion migration, have gained significant attention as potential candidates for both memory and neuromorphic applications. Among other perovskite-related materials, the memristive behavior of La₂NiO_4+δ (L2NO4)-based devices has been recently investigated using different electrodes and device configurations [1–4]. In the L2NO4 structure, interstitial oxygen serves as a negative charge defect, so its presence as a highly mobile ion in combination with a reactive TiN electrode leads to the creation of a TiN_xO_y interlayer at the metal/oxide interface due to the oxygen surplus in the L2NO4 films [4]. It was demonstrated that the TiN/L2NO4/Pt devices show non-volatile resistive switching (RS) with unique properties, such as a “soft-forming” step, which contrary to most filamentary devices does not require the application of a higher voltage to initialize the RS process and does not induce binary RS in the device. Thus, TiN/L2NO4/Pt devices demonstrate gradual analog RS and the ability of the conductance update under the application of the voltage pulses, i.e. long-term potentiation (LTP) and depression (LTD) when used as synaptic devices in neuromorphic systems [4]. It was also previously shown that the oxygen content in the L2NO4 film can be modified by post-deposition annealing in a wide range of oxygen stoichiometry, which can strongly influence the RS behavior [2].

In this work, the resistive switching behavior of the TiN/L2NO4/Pt devices is thoroughly investigated based on the thermally annealed L2NO4 films. We show that it is possible to control the interstitial oxygen content (δ) in the L2NO4 films by annealing using a reducing (Ar) or oxidizing (O₂) atmosphere, which allows for the tuning of the RS properties of the devices. The difference in the initial stoichiometry of the devices was confirmed by XRD and XANES analysis. We show that the annealing in the Ar atmosphere increases the initial resistance of the device, resulting in the filamentary behavior with the classical forming step. At the same time, the filamentary RS in the Ar annealed sample exhibits LTP/LTD with µs range pulses, allowing to achieve a larger memory window at the reduced energy consumption. On the contrary, thermal annealing in the O₂ atmosphere leads to the forming-free behavior with HRS and LRS resistance depending on the device area, which indicates that the interfacial type switching is dominant in the O₂ annealed device. However, the slower switching kinetics of this device leads to an increase in power consumption during the potentiation and depression of synapses. Nevertheless, both devices demonstrated the ability of LTP/LTD and STDP-based learning, which makes both devices promising candidates for neuromorphic applications for different types of architectures.

The technology for fabricating microelectrode arrays (MEAs) has existed since the 1970s and extracellular electrophysiological recordings become well established in neuroscience, drug screening and cardiology. The arrays allow probing physiological and pathological spiking activity of large ensembles of excitable cells, under highly controlled physico-chemical conditions and noninvasively, with high temporal resolution. A number of laboratories worldwide are nonetheless active to search for alternatives and ways to increase performance and capabilities of these devices. 

In this talk, I will first introduce and motivate the research field of neuroprosthetics and brain pacemakers, and why advances in the interfacing technologies and materials of MEAs are pivotal.
I will then present recent progress in the fabrication and in vitro experimental validation of MEAs with the best-reported combination of small size and low electrochemical impedance and finally review some earlier results combining nanostructured materials as scaffolding and interfaces.

With increasingly simpler fabrication processes and wide availability of nanomaterials providers, these steps are suggesting an exciting future for the study of neuronal networks in vitro and in vivo.

Vision restoration in patients with outer retinal degenerative diseases, such as Age-related Macular Degeneration and Retinitis Pigmentosa can be achieved by bypassing the degenerated photoreceptors and the electrical stimulation of the relatively well-preserved inner retina through electrode implants. Although current retinal prostheses have been shown to provide useful vision in blind patients, the obtained visual acuity and quality are still relatively low, because of inherent limitations of current retinal prosthetic technologies: long electrodes-neuron distance with increased current activation threshold and non-selective neural activation. We are developing a novel approach to overcome these limitations by a hybrid retinal composed of a high-density electrode array (pixel distance down to the cellular size of 10µm), where each electrode is coupled with a glutamatergic neuron to create a tight neuron-electrode coupling. Following implantation of the hybrid prosthesis, the glutamatergic neurons integrate and synapse with the host retinal circuits. Patterned electrical stimulation of these glutamatergic neurons by the electrodes modulates glutamate release onto the synapse with the host bipolar cells after which the remaining retinal circuitry is activated identically to natural vision while keeping high resolution and selective retinal cells activation. We have shown that hESC-derived photoreceptor precursors create glutamatergic synapses in-vitro; the cells can be activated by the electric field, as was shown with patch-clamp and calcium imaging.  Confocal imaging revealed that the cell created a tight-seal configuration within the micro-well device. Following implantation to the subretinal space of rats with retinal degeneration, the precursors survived for over a month; the cells show axonal elongation to the inner retina. This presentation will further discuss various challenges, design considerations and potential solutions and approaches toward the development of a hybrid retinal implant.

The ability to manipulate and direct neuronal growth has great implications in basic science and tissue engineering. Physical mechanical forces, contact guidance cues and chemical cues play key roles in neuronal morphogenesis and network formation. In this talk I will present our recent studies of 2D and 3D nanostructured scaffolds as platforms for controlling neuronal growth. We grow neurons on substrates patterned with nanotopographic cues of different shapes and materials and study the effects on neuronal geometry, dynamics and function. We compare neurons interfacing the nano-patterned substrates to neurons interfacing other neurons to reveal mechanisms translating interactions into neuronal growth behavior. In-vivo, neurons grow in a 3-dimensional (3D) extracellular matrix (ECM). Imitating the 3D environment resembling the in-vivo conditions is important for effective regeneration post trauma. We have chosen a collagen hydrogel system as the 3D ECM analog to best mimic the natural environment and develop methods to orient the collagen fibers and use them as leading cues for neurons [1]. Current efforts to implant these modified gels to bridge gaps in injured PNS models effectively will be presented.

We also used magnetic manipulations, via MNPS, as mediators to apply forces locally on neurons and their environment, as drug carriers and as a method to organize cells remotely. We transform cells into magnetic units by the uptake of magnetic nanoparticles [2,3], as well as by coating the cells with magnetic particles[4,5]. Our system opens new possibilities for improved neuronal interfaces and neural network analysis for long time periods.

References:

Nonvolatile memory (eNVM) is an essential component of microsystems in the era of internet of things (IoT). With the CMOS technology node scaling below 28 nm, back-end-of-line memories, such as phase change memory (PCM), resistive switching memory (RRAM) and spin-transfer torque magnetic memory (STTRAM) have attracted strong interest due to their compatibility with high-k/metal-gate (HKMG) process, high density and low cost. In addition to non-volatile storage, these devices can display multilevel operation combined with excellent scaling, which makes them extremely promising for neuromorphic computing.

This talk will present an overview of the NVM technology, including their storage principle, maturity, density and reliability. Focusing on PCM and RRAM technologies, I will illustrate and compare their properties and challenges for analog in-memory computing (AIMC) and brain-inspired neuromorphic systems. Subthreshold-operated PCM and RRAM with one-selector(one-resistor structure will also be discussed in the frame of improving the energy efficiency. The most critical roadblocks, such as multilvel precision, reliability and system-level complexity will be discussed.

Neuromorphic systems have experienced a rapid upswing in the last decade due to the increasing spread of machine learning systems, but also due to new technological possibilities through memristive devices that enable efficient hardware realizations of bio-inspired computing paradigms. In this talk, an auditory neuromorphic system will be presented that emulates the outstanding characteristics of the human sense of hearing. This includes a dynamic range of more than 120 dB sound pressure level (SPL), a frequency resolution of up to 0.1 %, an intensity discrimination of only 1 dB and adaptability at low sound levels in noisy environments. The neuromorphic system presented here emulates the properties of the sense of hearing by means of critical coupled oscillators. Therefore, a micro-electromechanical system- (MEMS) with feedback electronic is used to realize the sensory part, while networks based on memristive devices are used for subsequent information processing. Using this system as an example, important requirements for MEMS sensors and memristive devices will be discussed, and it will be shown how a new way of information processing beyond current approaches can open a new bio-inspired pathway toward the construction of cognitive electronics.

As Moore’s law slows down and memory-intensive tasks get prevalent, digital computing becomes increasingly capacity- and power-limited. In order to meet the requirement for increased computing capacity and efficiency in the post-Moore era, emerging computing architectures, such as in-memory computing and neuromorphic computing architectures based on memristors, have been extensively pursued and become an important candidate for new-generation non-von Neumann computers. Here, we report development of highly compact artificial neurons and synapses, especially that capture important neuronal and synaptic dynamics, as well as the construction of hardware systems based on such artificial elements. These devices and systems are of great significance for developing neuromorphic hardware with augmented information processing and learning capabilities. We report an optoelectronic synapse that is based on α-In2Se3 and has controllable temporal dynamics under electrical and optical stimuli. Tight coupling between ferroelectric and optoelectronic processes in the synapse can be used to realize heterosynaptic plasticity, with relaxation timescales that are tunable via light intensity or back-gate voltage. We use the synapses to create a multimode reservoir computing system with adjustable nonlinear transformation and multisensory fusion, which is demonstrated using a multimode handwritten digit recognition task and a QR code recognition task. We also realize a multiscale reservoir computing system via the tunable relaxation timescale of the α-In2Se3 synapse, which is tested using a temporal signal prediction task [1]. Moreover, we propose a highly efficient neuromorphic physiological signal processing system based on VO2 memristors [2]. The volatile and positive/negative symmetric threshold switching characteristics of VO2 memristors are leveraged to construct a sparse-spiking yet high-fidelity asynchronous spike encoder for physiological signals. Besides, the dynamical behavior of VO2 memristors is utilized in compact Leaky Integrate and Fire and Adaptive-LIF neurons, which are incorporated into a decision-making Long short-term memory Spiking Neural Network. The system demonstrates superior computing capabilities, needing only small-sized LSNNs to attain high accuracies of 95.83% and 99.79% in arrhythmia classification and epileptic seizure detection, respectively. This work highlights the potential of memristors in constructing efficient neuromorphic physiological signal processing systems and promoting next-generation human-machine interfaces.

Memristive devices based on the valence change mechanism (VCM) are promising candidates for emerging memory and neuromorphic applications. Due to their two-terminal structure and the possibility to be arranged in large arrays, one of their largest potential is seen in inference-based tasks such as the operation of neural networks once the training is finished.

VCM-type memristive systems can be further classified based on the area scaling of the device conductance states. In particular, area-independent, i.e. filamentary-switching, devices and area-type switching devices exist. Recent advances in the understanding of the different conductance mechanisms [1] and of the area-dependent switching variety in particular [2] have identified this second class of VCM-type memristive devices a new prospect for aforementioned inference tasks. Due to their area-scaling conductance values and the improved signal-to-noise ratio over filamentary-switching devices [3], they appear to be the superior choice in this context. In contrast, area-dependent devices frequently suffer from slower switching speeds at increased switching voltage and limited retention times compared to the filamentary class.

In this work, the two classes are compared with one another in the context of an inference-based task applied on two typical candidates, namely Pt/HfO₂/TiO_x/Ti cells for the filamentary and Pt/Al₂O₃/TiO_x/Pt devices for the area-dependent class. Their conduction mechanism, switching kinetics, noise and retention properties are determined experimentally and fed into the simulation of a single-layer neural network for classification of the MNIST dataset. A comprehensive comparison highlights advantages and drawbacks of the device choice. Finally, strategies for exploiting the respective advantages while mitigating the respective drawbacks are evaluated, and a synergetic approach is proposed.

References

[1]      

[2]      

[3]      

Emerging memristor-based neuromorphic computing architectures heavily rely on the proper control of intrinsic fluctuations in the building blocks. In most neural network applications, the improper device noise prevents fine control over synaptic weights encoded in memristive conductances, and dedicated denoising techniques are needed for improved bit resolution [1]. In contrast, other schemes, such as memristive Hopfield neural networks, rely on a probabilistic computational approach, where device noise can be harvested as a resource for efficient convergence to a global solution of the targeted problem [2,3].

Here, we present a thorough analysis of the 1/f-type noise properties of various memristive systems [4-6]. We demonstrate detailed noise tailoring considerations, i.e. the dependence of the steady state noise amplitudes on the material choice, the junction size, the transport mechanism, and the characteristic geometry of the active volume. In addition to the steady-state noise properties, we present a method for full-switching-cycle nonlinear noise spectroscopy. This approach traces the transformation of current-voltage nonlinearities into nonlinear noise spectra, providing a unique tool to identify the relevant source of fluctuations in the transport model. Furthermore, noise measurements over the entire switching cycles reveal voltage-manipulated noise contributions in the non-steady-state regime, i.e. a cycle-to-cycle redistribution of the fluctuators around the device bottleneck. This feature highlights the rather strong cycle-to-cycle variation of noise in apparently reproducible resistive switches, while also allowing the selection of optimized junction states with highly denoised characteristics. With these studies, we demonstrate the merits of advanced noise spectroscopy, which not only provides a fundamental understanding of the sources of fluctuations, but also lays the foundation for noise reduction strategies for high precision neuromorphic applications.

Reservoir computing (RC) is a machine learning framework initially conceived as an alternative to mitigate the expensive cost in training recurrent neural networks [1]. A RC system consists of a reservoir, a nonlinear dynamic system that maps the input in a high-dimensional space, and a readout layer, a single layer network that can be trained with simple methods, e.g. a linear regression. The fact that the reservoir does not require training, relaxes the hardware requirements for a physical realization, thus motivating the use of the RC as a powerful tool to exploit numerous physical systems as data processing units [2].

In this study, we use the RC approach to compute tasks with a chaotic oscillator, a modified version of the Murali-Lakhsmanan-Chua circuit [3]. More concretely, we included a Pt/HfO₂/TiN nonvolatile memristive device in the circuit. Such devices exhibit a physical phenomenon called resistive switching consisting in a nonvolatile change of resistance upon the application of voltage across its terminals. In the particular case of our device, the resistance increases (decreases) only after surpassing a positive (negative) threshold voltage level. When included in the circuit, the memristive device acts as a programmable resistor, providing a mechanism to tune the circuit's overall dynamics. The circuit is used as a reservoir, the first stage of data processing. The input data is masked in the amplitude of a periodic signal that induces the circuit to generate oscillations. Since the circuit has a single output port, the whole transient response is tracked with a sampling stage, where each sample constitutes a virtual reservoir output, and therefore a feature for training the readout layer. We demonstrate the feasibility of this scheme experimentally, with a physical realization of the circuit; only the training and inference of the readout layer are executed off-line. We tested the approach with two tasks, the computation of logic functions and pattern classification. The whole RC scheme is able to successfully classify the inputs due the nonlinearity of the circuit. Results indicate that the presence of chaotic sequences does not directly imply a disadvantage, and that they can contribute to the computation as well. Finally, we show that the nonvolatile memristive device acts as a knob to improve the classification of the inputs, when the computing performance appears to be suboptimal.

Resistive random-access memory (ReRAM) devices have significant potential for low power in-memory and neuromorphic computing. However, there are major challenges that need to be addressed to further progress in this field, particularly in terms of integration into complementary metal-oxide-semiconductor (CMOS) technology [1]. Moreover, the majority of the reported ReRAM devices exhibit low resistance causing sneak current and high-power consumption in circuit applications [2]. Therefore, they require a rectifier for the control of the low resistance state (LRS). In this work, we introduce a new CMOS-compatible self-rectifying bilayer W/GaO_x/HfO₂/Ti memristive device, where thin amorphous GaO_x (5 nm) and HfO₂ (4 nm) layers are deposited by atomic layer deposition at low temperature (250 °C). The I-V characteristics of the devices exhibit a hysteretic behavior with access to low resistance and high resistance states with an operating voltage of 3.3 V. Both the HRS and LRS values scale with the electrode area, suggesting a bulk/interfacial-governed switching mechanism. The amorphous GaO_x layer provides rectification (no need for a compliance current) due to its semiconducting nature that can be tuned with the deposition parameters [3]. A potentiation depression behavior is evidenced under an identical pulse scheme programming with a low operation voltage (V_write = 3.2 V). In this voltage range, good compatibility with CMOS devices is achieved. Low writing (~ 260 nA/µm²) and reading currents (~100- 600 pA/µm²) can enable low power applications. Moreover, the current levels can be tailored to the needs of a specific circuit by simply changing the electrode area. We exploit the analog behavior of this new bilayer GaO_x/HfO₂ memristive device to detect high frequency activity of neuronal cell culture. The device response under pre-recorded neuronal signals (spike trains) is studied and shows a potential for integration in the back-end-of-line of hybrid CMOS/microelectrode array (MEA) chips for real-time neuronal activity processing.

Ion-based neuromorphic devices have attracted wide interest as the basic building block of neuromorphic signal processing [1]. Here, we describe a multipore nanofluidic memristor with conical pores on a polymeric substrate that shows a wide range of ionic conduction properties, including current rectification [2]. These properties are based on the electrical interaction between the functionalized charges on the conical pore surface and the nanoconfined ionic solution [3]. The memristor shows a range of ionic conduction properties that can be controlled by the amplitude and frequency of the voltage signal, the salt type, the ionic concentration, and the solution pH. The multipore membrane allows reliable responses by maximizing current output through an ensemble of pores. Also, it does not demand difficult operational procedures because the physico-chemical signals involved here are usual in electrochemical systems and can be easily handled. This wide modulation of ionic conduction and rectification is highly desirable for implementing logic functions, memory, and transistor characteristics that are central to complex neuromorphic phenomena. Further, the chemical gating switched in current and polarity by means of the pH control provides additional functionality for chemical computation and neuromorphic applications including synaptic potentiation and depression.

Memristors are candidates for scaled-down brain-inspired neuromorphic circuits because of their simple two-terminal (2T) device geometry and in-memory computation capability which can overcome the power limitations of the von Neumann architecture. Crossbar circuits based on 2T memristors typically require an additional unit such as a transistor for individual node selection. Although highly effective, this approach significantly increases circuit footprint and manufacturing complexity. A memristive device with gate-tunable synaptic functionalities would not only integrate selection functionality at the cell level but could also lead to enriched on-demand learning schemes. Here, a three-terminal (3T) mixed-halide perovskite memristive device with gate-tunable synaptic functions operating at low potentials is demonstrated [1]. The device operation was controlled by both the drain (V_D) and gate (V_G) potentials, with an extended endurance of >2000 cycles and a state retention of >5000 s. Applying a voltage (V_set) of 20 V across the 50 μm channel switches its conductance from a high-resistance-state (HRS) to low-resistance-state (LRS). A memristive switching mechanism is proposed that is supported by current injection models through a Schottky barrier and Kelvin probe force microscopy data. The simultaneous application of a V_G potential is found to further modulate the channel conductance and reduce the operating V_set to 2 V, thus requiring a low electric field of 400 V/cm, which is by a factor of 50× less compared to state-of-the-art literature reports. Gate-tunable retention, endurance and synaptic functionalities were demonstrated, further highlighting the beneficial effect of V_G on device operation. By setting appropriate current compliance current, the devices can be operated in volatile I-V switching mode demonstrating extended endurance characteristics [2]. In this diffusive memristor mode, the devices exhibit pulse-amplitude and -frequency characteristics allowing linear conductivity modulation opening the path for the implementation of a leaky integrate-and -fire (LIF) neuron with light and gate tunable functions.

[1] Rogdakis, K.; Chatzimanolis, K.; Psaltakis, G.; Tzoganakis, N.; Tsikritzis, D.; Anthopoulos, T. D.; Kymakis, E Mixed-halide perovskite memristors with gate-tunable functions operating at low switching electric fields. Adv. Electron. Mater.2023, 2300424

[2] Rogdakis K., et al. Manuscript in preparation.

The ionic conductivity of halide perovskite is responsible for a memory effect that can be used in resistive memories. Here, the use of oxidized metal buffer layers are shown to control the working mechanism.¹ The electronic response turns from being dominated by halide vacancies migration to a response dominated by metal migration. Pre-oxidized electrochemically active metal buffer layers show improved performance in terms of stability and reduced operational voltages. Several metals are evaluated including some of those typically employed in metal oxide memristors such as Ag, Al, Au and Pt. The metal contacts are evaluated together with 3D and 2D perovskites.^2,3,4 Silver Iodide as a source of oxidized metal show superior performance to the metals since the oxidized silver can readily follow the electrical field without the need of an electroforming step. Overall, we provide solid understanding on the operational mechanism of halide perovskite memristors that has enabled increased stabilities approaching 10⁵ cycles with well separated states of current and further improvements expected.⁵

Brain-inspired computer architectures have emerged as promising solutions for enhancing energy efficiency of machine learning models such as artificial neuronal networks (ANNs), typically characterized by high energy consumptions when implemented with conventional CMOS processors. Among other emerging components for ANNs, Synaptic Electrolyte-Gated Transistors (EGTs) have emerged as promising solutions to overcome the Van Neumann bottleneck, which is the fundamental limitation in traditional digital computing systems caused by the separation of data memory and processing units. EGTs are essentially conventional Field-Effect Transistors where the gate dielectric is replaced by an electrolyte with mobile ions. These ions can accumulate or intercalate in the channel modifying its conductance in an analogic and permanent way, mimicking the behaviour of a natural synapsis and drastically reducing the energy consumptions. However, they are currently based on electrolytes that are intrinsically unstable and difficult to integrate such as ionic liquids or proton conducting polymers, which are obviously sensitive to humidity and temperature and present a poor compatibility with mainstream microelectronics fabrication processes.

In this work, we will present an all-solid-state oxygen-ion synaptic transistor based on a symmetric battery-like device in which oxygen-ions are transferred between two mixed ionic-electronic conducting (MIEC) perovskite oxide thin-films (i.e. channel and gate) through a solid-state electrolyte. Oxygen ions intrinsically present an enhanced stability and compatibility compared to other smaller ions such as Li+ or H+ but the sluggish oxygen diffusivity of solid-state materials has typically limited the application to liquid electrolytes. In order to address this important issue, a superior oxide-ion conducting thin film of Bi2V1‑xCuxO(11/2)‑x (BICUVOX) with highest oxygen conductivities at low temperature ever reported is used as low temperature electrolyte.[1] The device was manufactured by pulsed laser deposition (PLD) and conventional microfabrication techniques. The conductance of the channel is controlled via a gate bias, allowing for oxygen-ions intercalation between gate and channel as well as multistate operation in the millisecond range. In-situ electrical conductivity measurements at low temperatures (< 150 °C) reveal fast and remarkable low energy consumption. Due to its multi-states, the presented EGT reveals synaptic features such as short-term plasticity (STP) and paired pulse facilitation (PPF), which are known as fundamental properties of synaptic plasticity in biological neural networks. The proposed synaptic transistor has the potential to lead to a breakthrough in energy efficiency and processor performance in information technology.

Next-generation optoelectronics aims to achieve the transition to smart wearable and flexible devices that can communicate with each other and perform neuromorphic computing at the edge. These devices should be able to carry out their regular tasks with the help of energy-efficient in-memory calculations. In this work, we fabricate optoelectronic memristors based on halide-perovskite microwires. The CsPbBr₃ halide-perovskite microwires are fabricated on a flexible polymer substrate and integrated with a thin film electrode made of single-walled carbon nanotubes in a lateral geometry. By applying hybrid optoelectrical stimuli, we have shown that our device can perform regular photodetection functions complemented by synaptic functionality. Importantly, we have demonstrated that our device exhibits frequency-dependent bidirectional modification of synaptic weight with a sliding threshold similar to biologically plausible Bienenstock-Cooper-Munro learning. We explain this complex behavior by competing capacitive and inductive branches of equivalent electrical circuit. Our work unveils the opportunity for the development of hybrid organic-inorganic artificial visual systems based on halide-perovskite single-crystals.

Memristive crossbar arrays show promise as non-von Neumann computing technologies, bringing sophisticated neural network processing to the edge and facilitating real-world online learning. However, their deployment for real-world learning problems faces challenges such as non-linearities in conductance updates, variations during operation, fabrication mismatch, conductance drift, and the realities of gradient descent training.
This talk will present methods to pre-train neural networks to be largely insensitive to these non-idealities during learning tasks. These methods rely on a phenomenological model of the device, obtainable experimentally, and bi-level optimization. We showcase this effect through meta-learning and a differentiable model of conductance updates on few-shot learning tasks. Since pre-training is a necessary procedure for any online learning scenario at the edge, our results may pave the way for real-world applications of memristive devices without significant adaptation overhead.
Furthermore, by considering the programming of memristive devices as a learning problem in its own right, we demonstrate that the developed methods can accelerate existing write-verify techniques.

The exploration of magnetic property manipulation through ionic motion in ferromagnetic/oxide structures has emerged as a promising avenue for non-volatile control of magnetism in spintronics devices. This concept unlocks new possibilities, such as the development of reconfigurable multistate memories and the incorporation of cumulative gate effects. Inspired by memristor technologies, oxygen-based magneto-ionics stands out as a cutting-edge approach, providing an advanced framework for influencing magnetic properties through ionics.

In this presentation I will provide an overview of this exciting field, and a description of the underlying physical-chemical mechanisms at play. I will also show in detail our work [1-4] on CoFeB/oxide systems, particularly Ta/CoFeB/HfO₂, where ionic gating induces the migration of oxygen-rich species within the stack. This migration results in distinct magneto-ionic regimes characterized by varying degrees of oxygen content, thereby allowing precise control over the oxidation state of magnetic layers—ranging from under-oxidized to optimally oxidized to over-oxidized. Notably, these different magneto-ionic regimes give rise to diverse spin-reorientation transitions, showing varying degrees of reversibility. In addtion, I will also discuss our recent efforts in designing artificial synaptic elements using magneto-ionic devices based on CoFeB/HfO₂. 

The expression “neuromorphic computing” strikingly shows that the contemporary scientific narrative compares brains to computers [1]. Computer architectures are created by designers and deterministically assembled, whereas biological neural networks are based on self-organization so that structure and function co-evolve. Both process data using a huge number of simple equivalent components linked by a complex physical pattern of connections, however it remains quite undefined to what extent the use of the brain-computer metaphor can be useful for the design of computers that perform as brains.

Self-assembled nanoparticle or nanowire networks have recently come under the spotlight as systems able to obtain brain-like data processing performances by exploiting the memristive character and the wiring of the junctions connecting the nanostructured network building blocks [2]. The role of the wiring in biological and artificial systems has been originally recognized by Rosenblatt with the Perceptron model [3], a more complex model was proposed by Hopfield who pointed out the importance of the non-linearity of the input-output relationship and of the use of nonlinear logical operations [4]. These aspects are crucial for the design and operation of data processing networks of nanoobjects.

Recently we demonstrated that nanostructured Au films, fabricated by the assembling of gold clusters produced in the gas phase, have non-linear and non-local electric conduction properties caused by the extremely high density of grain boundaries and the resulting complex arrangement of nanojunctions [5]. Starting from the characterization of this system, we proposed and formalized a generalization of the Perceptron model to describe a classification device based on a network of interacting units where the input weights are non-linearly dependent [6]. This model, called ‘‘Receptron’’, provides substantial advantages compared to the Perceptron as, for example, the solution of non-linearly separable Boolean functions with a single device [6].

Here we present and discuss the practical application of the Receptron model to the realization of electronic components for the classification of Boolean function without previous training in view of the fabrication of arithmetic logic unit circuits.

We will also show that the Receptron model can be used for the implementation of an all-optical device exploiting the non-linearity of optical speckle fields produced by a solid scatterer. A single layer optical Receptron network can efficiently be used for pattern recognition without previous training. The optical implementation of the receptron scheme opens the way for the fabrication of a completely new class of optical devices for neuromorphic data processing based on a very simple hardware.

The replication of neural information processing in electrical devices has been extensively studied over the years. The paradigm of parallel computing, which allows information to be simultaneously detected, processed, and stored, is required for numerous applications in many fields. In the case of brain-computer interfaces, another important requirement is the suitability of the device for communication with cells. Organic electrochemical transistors (OECTs) based on PEDOT:PSS are used for this purpose due to their ionic-to-electronic signal transduction and biocompatibility [1]. Many works have demonstrated the reproduction of neural plasticity mechanisms, such as short-term facilitation and long-term potentiation. In each device, the physical mechanism of transduction may be different, but it is known that the electrolyte plays a key role in the functioning of these devices, as it provides the ions responsible for the chemical transmission of information. Focusing on long-term memory, this can be reproduced in the OECTs with the oxidation of the neurotransmitter, as in the case of the biohybrid synapse [2]. It is crucial to understand the influence of the material chemistry and the electrolyte composition on the memory effect of the device, as long-term modulation is based on a change in the ionic balance between the electrolyte and the organic polymer.

This electrolyte-dependency plasticity will be discussed as it should be considered when the OECT is used in a biological environment in which many molecules of a different nature are present in addition to neurotransmitters. Furthermore, I will discuss how conjugated polymers can be engineered with azopolymers (opto-sensitive polymers which switch from cis to trans conformation upon certain light exposure) to feature diverse optoelectronic short- and long-term plasticity, enabling the use of such platforms as neurohybrid devices as building blocks of retina-inspired devices

References
1 Bernard et al. (2007). In: Advanced Functional Materials 17.17, pp. 3538–3544.
2 Keene, Scott T et al. (2020). In: Nature Materials 19.9, pp. 969–973.

Analog tunable memristors are widely utilized as artificial synapses in various neural network applications. However, exploiting the dynamical aspects of their conductance change to implement active neurons is still in its infancy, awaiting the realization of efficient neural signal recognition functionalities. Here we experimentally demonstrate an artificial neural information processing unit that can detect a temporal pattern in a very noisy environment, fire an output spike upon successful detection and reset itself in a fully unsupervised, autonomous manner [1]. This circuit relies on the dynamical operation of only two memristive blocks: a non-volatile Ta2O5 device and a volatile VO2 unit. A fading functionality with exponentially tunable memory time constant enables adaptive operation dynamics, which can be tailored for the targeted temporal pattern recognition task. In the trained circuit false input patterns only induce short-term variations. In contrast, the desired signal activates long-term memory operation of the non-volatile component, which triggers a firing output of the volatile block. Possible applications of the presented scheme in larger-scale reservoir computing architectures are also discussed.

Optoelectronic memristors open new opportunities for neurosynaptic devices and optoelectronic systems.[1] The phenomenon of photoluminescence present in many semiconductors gives us a possibility to learn about charge recombination using rather simple optical spectroscopy methods. In this talk, I will describe a novel technique of automotive mapping of photoluminescence quantum yield leading to a PLQY (f,P) map, which is also called a “Horse map”.[2,3] This map presents PLQY of a material in the space of the laser pulse energy (P) and the laser repletion rate (f) where the scanning of P and f occurs over many orders of magnitude in a pre-designed programmable manner.

The “Horse map” is a fingerprint of both charge carrier dynamics occurring at timescales from nanoseconds to microseconds and of the defect dynamics happening at time scales from milliseconds to hours. For the latter the automatization of the mapping is crucial since the defect evolution occurs during the experiment itself.[3] It means that the methodology is applicable for materials which change their properties over time, like, for example, a material of memristor under repetitive writing/reading cycles.

I will further unveil how PLQY(f,P) mapping allows to extract the concentration of defect states and their nature. I will describe the so called “observer effect” in metal-halide perovskites which is the consequence of the defect dynamics.[3]

I envision that further development of PLQY mapping over a multi-parameter space not limited to f and P, for example using hybrid optoelectrical stimuli, will be desired for the optoelectronic memristor community to rationalize charge carrier dynamics under real device operation conditions.