Search

Reservoir computing (RC)-based neuromorphic applications exhibit extremely low power consumption, thus challenging the use of deep neural networks in terms of both consumption requirements and integration density. Under this perspective, this work focuses on the basic principles of RC systems. The ability of self-selective conductive-bridging random access memory devices to operate in two modes, namely, volatile and non-volatile, by regulating the applied voltage is first presented. We then investigate the relaxation time of these devices as a function of the applied amplitude and pulse duration, a critical step in determining the desired non-linearity by the reservoir. Moreover, we present an in-depth study of the impact of selecting the appropriate pulse-stream and its final effects on the total power consumption and recognition accuracy in a handwritten digit recognition application from the National Institute of Standards and Technology dataset. Finally, we conclude at the optimal pulse-stream of 3-bit, through the minimization of two cost criteria, with the total power remaining at 287 µW and simultaneously achieving 82.58% recognition accuracy upon the test set.

The development of disruptive artificial neural networks (ANNs) endowed with brain-inspired neuromorphic capabilities is emerging as a promising solution to deal with the challenges of the artificial intelligence era. The fabrication of robust and accurate ANNs is strongly associated with the design of new electronic devices. The intriguing properties of memristors render them suitable as building blocks within ANNs. However, the impact of the operating electrodes on the dynamics of the switching process and the relaxation effect remains elusive. It is, thus, apparent that a deep understanding of the underlying electrochemical metallization mechanism that affects the formation of the conductive filament is of great importance. Along these lines, in this work, the impact of various materials as inert electrodes (Pt NPs, ITO, n++ Si, TiN, and W) on tuning the switching mode of low power SiO2-based conductive bridge random access memory devices was systematically investigated. A comprehensive model was applied to interpret the threshold and bipolar switching patterns and shed light on the respective physical mechanisms. The model incorporated the different coefficients of thermal conductivity of the various materials and attempted to associate them with the Soret coefficient and the activation energy of thermophoresis to interpret the experimental outcomes. Our work provides valuable insight for the realization of memristive devices with tunable properties, which can be directly leveraged for implementing a variety of neuromorphic functionalities, such as synaptic plasticity and spike generation. [ABSTRACT FROM AUTHOR]

The quick growth of information technology has necessitated the need for developing novel electronic devices capable of performing novel neuromorphic computations with low power consumption and a high degree of accuracy. In order to achieve this goal, it is of vital importance to devise artificial neural networks with inherent capabilities of emulating various synaptic properties that play a key role in the learning procedures. Along these lines, we report here the direct impact of a dense layer of Pt nanoparticles that plays the role of the bottom electrode, on the manifestation of the bipolar switching effect within SiO2-based conductive bridge memories. Valuable insights regarding the influence of the thermal conductivity value of the bottom electrode on the conducting filament growth mechanism are provided through the application of a numerical model. The implementation of an intermediate switching transition slope during the SET transition permits the emulation of various artificial synaptic functionalities, such as short-term plasticity, including paired-pulsed facilitation and paired-pulse depression, long-term plasticity and four different types of spike-dependent plasticity. Our approach provides valuable insights toward the development of multifunctional synaptic elements that operate with low power consumption and exhibit biological-like behavior.

Reservoir computing (RC)-based neuromorphic applications exhibit extremely low power consumption, thus challenging the use of deep neural networks in terms of both consumption requirements and integration density. Under this perspective, this work focuses on the basic principles of RC systems. The ability of self-selective conductive-bridging random access memory devices to operate in two modes, namely, volatile and non-volatile, by regulating the applied voltage is first presented. We then investigate the relaxation time of these devices as a function of the applied amplitude and pulse duration, a critical step in determining the desired non-linearity by the reservoir. Moreover, we present an in-depth study of the impact of selecting the appropriate pulse-stream and its final effects on the total power consumption and recognition accuracy in a handwritten digit recognition application from the National Institute of Standards and Technology dataset. Finally, we conclude at the optimal pulse-stream of 3-bit, through the minimization of two cost criteria, with the total power remaining at 287 µW and simultaneously achieving 82.58% recognition accuracy upon the test set. [ABSTRACT FROM AUTHOR]

The resistive switching characteristics of forming free TiO2–x/TiO2–y memory devices containing Pt nanocrystals (NCs) beneath the top electrode were systematically investigated through experiments and numerical simulation insights. By embedding Pt nanocrystals, we have the possibility to narrow down the possible locations where the switching effect will evolve and thus significantly improve the inherent variability of the devices. Besides, the deployment of bilayer structures can tune the resistance levels, since the presence of the layer with low oxygen content (TiO2–y) acts practically as series resistance, limiting the operating currents and at the same time forcing the switching effect to evolve in the layer with the higher oxygen content (TiO2–z). A numerical model is implemented, in order to shed light into the origin of the SET/RESET transitions and illustrate the direct impact of NCs on the conducting filament (CF) shape and distribution of oxygen vacancies. It is demonstrated that a higher density of oxygen vacancies is created in the vicinity of NCs, which can directly impact the operating current values and the uniformity of the switching characteristics. The presence of NCs also facilitates the reduction of the operating voltages (∼3 V), and, as a result, it significantly improves power consumption, without sacrificing the switching ratio (∼103), temporal/spatial variability (σ/μ < 0.2), and pulse endurance (108 cycles) characteristics of our memory cells. Evidence about the impact of the NCs position within the material configuration are also presented. The direct impact of Pt NCs on the depression and potentiation characteristics of the synaptic weight denotes similarly the huge applicability of our approach to tune a wide range of resistive switching properties. [ABSTRACT FROM AUTHOR]

The fabrication of low-power and scalable electronic devices that will have the ability to emulate the properties of the biological nociceptors is of great importance for the development of humanoid robots. Along these lines, in this work, an artificial nociceptive element composed of a SiO2/VOx-based bilayer configuration and a dense layer of Pt nanoparticles (NPs) as a bottom electrode is proposed. Interestingly, the device operates only under the threshold switching mode with the switching voltage as low as ∼220 mV and a huge switching ratio of 107. A systematic analysis of the impact of the bilayer configuration and the existence of the Pt NPs on the total memory performance is also provided, while a comprehensive numerical model is introduced to highlight the crucial role of the electrode material on the local temperature distribution and its influence on the memristive effect. On top of that, the proposed structure can imitate the normal, relaxation, and sensitization states of the nociceptors with about 0.3 pJ energy per spike. These enhanced properties are ascribed to the self-rupture of the Ag-based conducting filament, whereas valuable insights into the impact of the local temperature distribution on the switching dynamics are provided. [ABSTRACT FROM AUTHOR]

Although multilevel capability is probably the most important property of resistive random access memory (RRAM) technology, it is vulnerable to reliability issues due to the stochastic nature of conducting filament (CF) creation. As a result, the various resistance states cannot be clearly distinguished, which leads to memory capacity failure. In this work, due to the gradual resistance switching pattern of TiO2-x-based RRAM devices, we demonstrate at least six resistance states with distinct memory margin and promising temporal variability. It is shown that the formation of small CFs with high density of oxygen vacancies enhances the uniformity of the switching characteristics in spite of the random nature of the switching effect. Insight into the origin of the gradual resistance modulation mechanisms is gained by the application of a trapassisted-tunneling model together with numerical simulations of the filament formation physical processes. [ABSTRACT FROM AUTHOR]

A deep understanding of the underlying resistive switching mechanism for the implementation of volatile memristive properties is regarded as of great importance for enhancing their performance. Along these lines, a 2-D dynamical model is introduced to interpret the whole memristive pattern within the bilayer configuration, as well as the crucial of the dense layer of the Pt nanoparticles (NPs) on the local thermal distribution. Moreover, the probabilistic leaky-integrate-and-fire (LIF) neuron properties were simulated by considering a simple $\textit {RC}$ circuit in order to perform Bayesian extrapolation within a spiking neural network. A classification application is consequently demonstrated by using the liver tumor dataset. The advantageous capabilities of the stochastic-based spike neural networks (SNNs) are highlighted in striking contrast with the conventional artificial neural networks (ANNs), as well as the deterministic-based SNNs, in terms of prediction accuracy and power consumption. [ABSTRACT FROM AUTHOR]

The development of low-power neurons with an intrinsic degree of stochasticity is considered quite important for the emulation of the respective probabilistic procedures that take place within the biological neural networks. While the majority of the reported artificial neuronal configurations relies on the demonstration of well-defined spiking patterns or the realization of complicated neuromorphic properties, there is an urgent need to devise biological-like neurons with arbitrary firing capabilities. Along these lines, we present a novel threshold switching memristor from SiO2-based bilayer conductive-bridge resistive switching memory (CBRAM) and Pt nanoparticles (NPs) used as a bottom electrode (BE) in order to implement robust stochastic neuron activity accompanied with steep transition slope [~0.75 mV/dec(A)], large OFF-state (~ $10^{11} \Omega $), low switching voltage (~200 mV), and ultrafast turn-on speed (~50 ns). Moreover, probabilistic leaky-integrate-and-fire (LIF) neuron properties were attained by employing a simple RC circuit as well as tunable integrate-and-fire (IF) properties from stand-alone threshold switching elements, which are of great significance for the development of artificial spiking neuromorphic networks. [ABSTRACT FROM AUTHOR]

The operating principle of resistive random access memories (RRAMs) relies on the distribution of ionic species and their influence on the electron transport. Taking into account that formation and annihilation of conducting filaments (CFs) is the driving mechanism for the switching effect, it is very important to control the regions where these filaments will evolve. Thus, homolayers of titanium oxide with different oxygen contents were fabricated in order to tune the local electrical and thermal properties of the CFs and narrow down the potential percolation paths. We show that the oxygen content in the top layer of the TiO2-x/TiO2-y bilayer memristors can directly influence the morphology of the layers which affect the diffusion barrier and consequently the diffusivity and drift velocity of oxygen vacancies, yielding in important enhancement of switching characteristics, in terms of spatial uniformity (σ/μ<0.2), enlarged switching ratio (~104), and synaptic learning. In order to address the experimental data, a physical model was applied, divulging the crucial role of temperature, electric potential and oxygen vacancy density on the switching effect and offering physical insights to the SET/RESET transitions and the analog switching. The forming free nature of all the devices in conjunction with the self-rectifying behavior, should also be regarded as important assets towards RRAM device optimization. [ABSTRACT FROM AUTHOR]

In this work, we demonstrate that TiO2-x based Resistive Random Access Memory devices can function without an initial electroforming process and a wide range of switching ratios could be achieved by controlling the oxygen content, the compliance current, the sweep bias amplitude, and the width of the voltage pulse applied on the memory cell. The influence of deposition ambient and more particularly of oxygen flux during thin film sputtering at room temperature to the resistive properties of titanium oxide will be discussed in detail. By controlling the density of oxygen vacancies into the dielectric matrix, we can also improve the repeatability and the operation of the device, in terms of distribution of the SET/RESET voltages. We propose that ultra high density of vacancies deteriorate the switching phenomenon, whereas high vacancy density results in better switching behavior. Moreover, we conclude that the oxygen vacancies density and distribution have a direct impact on the conducting filament diameter, in terms of sensitivity of the conducting paths (high OFF/ON ratio). By increasing the oxygen content, we reduce the size of vacancy based filaments, resulting in a more stable operation of our device. In addition, manipulation of population of oxygen ions into the Ti top electrode enables the creation of multilevel switching states. Switching speed, endurance, and retention performance reveals the excellent functionality of our device as a non-volatile memory element and conduction mechanism analysis demonstrates the manifestation of Poole-Frenkel emission in conjunction with trap-assisted tunneling, which is also deployed in order to interpret the gradual increase of current during SET process. [ABSTRACT FROM AUTHOR]

The quick growth of information technology has necessitated the need for developing novel electronic devices capable of performing novel neuromorphic computations with low power consumption and a high degree of accuracy. In order to achieve this goal, it is of vital importance to devise artificial neural networks with inherent capabilities of emulating various synaptic properties that play a key role in the learning procedures. Along these lines, we report here the direct impact of a dense layer of Pt nanoparticles that plays the role of the bottom electrode, on the manifestation of the bipolar switching effect within SiO 2 -based conductive bridge memories. Valuable insights regarding the influence of the thermal conductivity value of the bottom electrode on the conducting filament growth mechanism are provided through the application of a numerical model. The implementation of an intermediate switching transition slope during the SET transition permits the emulation of various artificial synaptic functionalities, such as short-term plasticity, including paired-pulsed facilitation and paired-pulse depression, long-term plasticity and four different types of spike-dependent plasticity. Our approach provides valuable insights toward the development of multifunctional synaptic elements that operate with low power consumption and exhibit biological-like behavior.

The threshold switching effect is considered of outmost importance for a variety of applications ranging from the reliable operation of crossbar architectures to emulating neuromorphic properties with artificial neural networks. This property is strongly believed to be associated with the rich inherit dynamics of a metallic conductive filament (CF) formation and its respective relaxation processes. Understanding the origin of these dynamics is very important in order to control the degree of volatility and design novel electronic devices. Here, we present a synergistic numerical and experimental approach in order to deal with that issue. The distribution of relaxation time is addressed through time-resolved pulse measurements whereas the entire switching behavior is modeled through a 2D dynamical model by taking into account the destructive interference of the drift/diffusion transport mechanisms and the Soret diffusion flux due to the intense local Joule heating. The proposed mechanism interprets successfully both the threshold to bipolar switching transition as well as the self-rectifying effects in SiO 2 -based memories. The model incorporates the effect of electrode materials on the switching pattern and provides a different perception of the ionic transport processes, shading light into the ultra-small lifetimes of the CF and explaining the different behavior of the silver or copper active materials in a conductive bridge random access memory architecture.