A Hybrid CMOS-Memristor Spiking Neural Network Supporting Multiple Learning Rules

Artificial intelligence (AI) is changing the way computing is performed to cope with real-world, ill-defined tasks for which traditional algorithms fail. AI requires significant memory access, thus running into the von Neumann bottleneck when implemented in standard computing platforms. In this respect, low-latency energy-efficient in-memory computing can be achieved by exploiting emerging memristive devices, given their ability to emulate synaptic plasticity, which provides a path to design large-scale brain-inspired spiking neural networks (SNNs). Several plasticity rules have been described in the brain and their coexistence in the same network largely expands the computational capabilities of a given circuit. In this work, starting from the electrical characterization and modeling of the memristor device, we propose a neuro-synaptic architecture that co-integrates in a unique platform with a single type of synaptic device to implement two distinct learning rules, namely, the spike-timing-dependent plasticity (STDP) and the Bienenstock-Cooper-Munro (BCM). This architecture, by exploiting the aforementioned learning rules, successfully addressed two different tasks of unsupervised learning.

Hardware implementation of a true random number generator integrating a hexagonal boron nitride memristor with a commercial microcontroller

The development of the internet-of-things requires cheap, light, small and reliable true random number generator (TRNG) circuits to encrypt the data-generated by objects or humans-before transmitting them. However, all current solutions consume too much power and require a relatively large battery, hindering the integration of TRNG circuits on most objects. Here we fabricated a TRNG circuit by exploiting stable random telegraph noise (RTN) current signals produced by memristors made of two-dimensional (2D) multi-layered hexagonal boron nitride (h-BN) grown by chemical vapor deposition and coupled with inkjet-printed Ag electrodes. When biased at small constant voltages (≤70 mV), the Ag/h-BN/Ag memristors exhibit RTN signals with very low power consumption (∼5.25 nW) and a relatively high current on/off ratio (∼2) for long periods (>1 hour). We constructed TRNG circuits connecting an h-BN memristor to a small, light and cheap commercial microcontroller, producing a highly-stochastic, high-throughput signal (up to 7.8 Mbit s^-1) even if the RTN at the input gets interrupted for long times up to 20 s, and if the stochasticity of the RTN signal is reduced. Our study presents the first full hardware implementation of 2D-material-based TRNGs, enabled by the unique stability and figures of merit of the RTN signals in h-BN based memristors.

Emergence of associative learning in a neuromorphic inference network

OBJECTIVE

In the theoretical framework of predictive coding and active inference, the brain can be viewed as instantiating a rich generative model of the world that predicts incoming sensory data while continuously updating its parameters via minimization of prediction errors. While this theory has been successfully applied to cognitive processes - by modelling the activity of functional neural networks at a mesoscopic scale - the validity of the approach when modelling neurons as an ensemble of inferring agents, in a biologically plausible architecture, remained to be explored.

APPROACH

We modelled a simplified cerebellar circuit with individual neurons acting as Bayesian agents to simulate the classical delayed eyeblink conditioning protocol. Neurons and synapses adjusted their activity to minimize their prediction error, which was used as the network cost function. This cerebellar network was then implemented in hardware by replicating digital neuronal elements via a low-power microcontroller.

MAIN RESULTS

Persistent changes of synaptic strength - that mirrored neurophysiological observations - emerged via local (neurocentric) prediction error minimization, leading to the expression of associative learning. The same paradigm was effectively emulated in low-power hardware showing remarkably efficient performance compared to conventional neuromorphic architectures.

SIGNIFICANCE

These findings show that: i) an ensemble of free energy minimizing neurons - organized in a biological plausible architecture - can recapitulate functional self-organization observed in nature, such as associative plasticity, and ii) a neuromorphic network of inference units can learn unsupervised tasks without embedding predefined learning rules in the circuit, thus providing a potential avenue to a novel form of brain-inspired artificial intelligence.

Standards for the Characterization of Endurance in Resistive Switching Devices

Resistive switching (RS) devices are emerging electronic components that could have applications in multiple types of integrated circuits, including electronic memories, true random number generators, radiofrequency switches, neuromorphic vision sensors, and artificial neural networks. The main factor hindering the massive employment of RS devices in commercial circuits is related to variability and reliability issues, which are usually evaluated through switching endurance tests. However, we note that most studies that claimed high endurances >10⁶ cycles were based on resistance versus cycle plots that contain very few data points (in many cases even <20), and which are collected in only one device. We recommend not to use such a characterization method because it is highly inaccurate and unreliable (i.e., it cannot reliably demonstrate that the device effectively switches in every cycle and it ignores cycle-to-cycle and device-to-device variability). This has created a blurry vision of the real performance of RS devices and in many cases has exaggerated their potential. This article proposes and describes a method for the correct characterization of switching endurance in RS devices; this method aims to construct endurance plots showing one data point per cycle and resistive state and combine data from multiple devices. Adopting this recommended method should result in more reliable literature in the field of RS technologies, which should accelerate their integration in commercial products.

Advanced Data Encryption ​using 2D Materials

Advanced data encryption requires the use of true random number generators (TRNGs) to produce unpredictable sequences of bits. TRNG circuits with high degree of randomness and low power consumption may be fabricated by using the random telegraph noise (RTN) current signals produced by polarized metal/insulator/metal (MIM) devices as entropy source. However, the RTN signals produced by MIM devices made of traditional insulators, i.e., transition metal oxides like HfO₂ and Al₂ O₃ , are not stable enough due to the formation and lateral expansion of defect clusters, resulting in undesired current fluctuations and the disappearance of the RTN effect. Here, the fabrication of highly stable TRNG circuits with low power consumption, high degree of randomness (even for a long string of 2²⁴  - 1 bits), and high throughput of 1 Mbit s^-1 by using MIM devices made of multilayer hexagonal boron nitride (h-BN) is shown. Their application is also demonstrated to produce one-time passwords, which is ideal for the internet-of-everything. The superior stability of the h-BN-based TRNG is related to the presence of few-atoms-wide defects embedded within the layered and crystalline structure of the h-BN stack, which produces a confinement effect that avoids their lateral expansion and results in stable operation.

On the Modeling of the Donor/Acceptor Compensation Ratio in Carbon-Doped GaN to Univocally Reproduce Breakdown Voltage and Current Collapse in Lateral GaN Power HEMTs

The intentional doping of lateral GaN power high electron mobility transistors (HEMTs) with carbon (C) impurities is a common technique to reduce buffer conductivity and increase breakdown voltage. Due to the introduction of trap levels in the GaN bandgap, it is well known that these impurities give rise to dispersion, leading to the so-called “current collapse” as a collateral effect. Moreover, first-principles calculations and experimental evidence point out that C introduces trap levels of both acceptor and donor types. Here, we report on the modeling of the donor/acceptor compensation ratio (CR), that is, the ratio between the density of donors and acceptors associated with C doping, to consistently and univocally reproduce experimental breakdown voltage (V_BD) and current-collapse magnitude (ΔI_CC). By means of calibrated numerical device simulations, we confirm that ΔI_CC is controlled by the effective trap concentration (i.e., the difference between the acceptor and donor densities), but we show that it is the total trap concentration (i.e., the sum of acceptor and donor densities) that determines V_BD, such that a significant CR of at least 50% (depending on the technology) must be assumed to explain both phenomena quantitatively. The results presented in this work contribute to clarifying several previous reports, and are helpful to device engineers interested in modeling C-doped lateral GaN power HEMTs.

Electroforming in Metal-Oxide Memristive Synapses

Memristors have shown an extraordinary potential to emulate the plastic and dynamic electrical behaviors of biological synapses and have been already used to construct neuromorphic systems with in-memory computing and unsupervised learning capabilities; moreover, the small size and simple fabrication process of memristors make them ideal candidates for ultradense configurations. So far, the properties of memristive electronic synapses (i.e., potentiation/depression, relaxation, linearity) have been extensively analyzed by several groups. However, the dynamics of electroforming in memristive devices, which defines the position, size, shape, and chemical composition of the conductive nanofilaments across the device, has not been analyzed in depth. By applying ramped voltage stress (RVS), constant voltage stress (CVS), and pulsed voltage stress (PVS), we found that electroforming is highly affected by the biasing methods applied. We also found that the technique used to deposit the oxide, the chemical composition of the adjacent metal electrodes, and the polarity of the electrical stimuli applied have important effects on the dynamics of the electroforming process and in subsequent post-electroforming bipolar resistive switching. This work should be of interest to designers of memristive neuromorphic systems and could open the door for the implementation of new bioinspired functionalities into memristive neuromorphic systems.