Neuromorphic Hardware Research Articles

Artificial Synapse with High Weight-Updating Performance Based on Charge-Trapping Mechanism.

Artificial synapses, basic units of neuromorphic hardware, have been studied to emulate synaptic dynamics, which are beneficial for realizing high-quality neural networks. Currently, two-dimensional (2D) material heterojunction structures are widely used in the study of artificial synapses; however, their dynamic weight-updating characteristics are restricted owing to their high nonlinearity and low symmetricity. In this study, we treated h-BN with oxygen plasma to form a charge-trapping layer (CTL), and we prepared 2D ReS2/CTL/h-BN heterojunction synapses. The device achieves a large memory window and excellent synaptic performance and simulates the adaptive behavior of the human eye through the synergistic modulation of the optoelectronic double pulse. The mechanism of the effect of trap states in CTL on the weight-updating performance was analyzed, and the device was further optimized. In the long-term potentiation/depression (LTP/D) weight-updating characteristic of the device, the nonlinearity was reduced to 0.63 and symmetricity reached 41.25, which is superior to most similar devices reported to date. Therefore, this research provides insights into improving the LTP/D weight-updating performance of synaptic devices.

High-Performance Edge-Line Contact Memristors with In-Plane Solid-Liquid-Solid Grown Silicon Nanowires for Probabilistic Neuromorphic Computing.

Memristors have garnered increasing attention in neuromorphic computing hardware due to their resistive switching characteristics. However, achieving uniformity across devices and further miniaturization for large-scale arrays remain critical challenges. In this study, we demonstrate the scalable production of highly uniform, quasi-one-dimensional diffusive memristors based on heavily doped n-type silicon nanowires (SiNWs) with diameters as small as ∼50 nm, fabricated via in-plane solid-liquid-solid (IPSLS) growth technology. The edge-line contact structural design improves the control of nucleation sites and the size of conductive filaments (CFs) in Ag/SiO2/n-SiNW memristors. These devices exhibit excellent self-compliance threshold switching characteristics, including a low operating voltage (∼0.8 V) with a standard deviation of 0.073 V, low leakage current (1 pA), high switching ratio (>107), ultrafast switching speed (∼8 ns), and extremely low switching energy (47.2 fJ per operation). Additionally, we developed neurons with tunable sigmoidal probabilistic activation functions, demonstrating high uniformity across different devices. These neurons achieved an accuracy of 96.2% in binary tumor classification tasks, underscoring the potential of IPSLS-fabricated SiNWs for advanced neuromorphic computing hardware. This work highlights the effectiveness of SiNW-based memristors in addressing challenges in neuromorphic hardware design and their potential for large-scale integration.

A self-training spiking superconducting neuromorphic architecture

A self-training spiking superconducting neuromorphic architecture

Accelerating spiking neural networks with parallelizable leaky integrate-and-fire neurons*

Abstract Spiking neural networks (SNNs) express higher biological plausibility and excel at learning spatiotemporal features while consuming less energy than conventional artificial neural networks, particularly on neuromorphic hardware. The leaky integrate-and-fire (LIF) neuron stands out as one of the most widely used spiking neurons in deep learning. However, its sequential information processing leads to slow training on lengthy sequences, presenting a critical challenge for real-world applications that rely on extensive datasets. This paper introduces the parallelizable LIF (ParaLIF) neuron, which accelerates SNNs by parallelizing their simulation over time, for both feedforward and recurrent architectures. Compared to LIF in neuromorphic speech, image and gesture classification tasks, ParaLIF demonstrates speeds up to 200 times faster and, on average, achieves greater accuracy with similar sparsity. When integrated into state-of-the-art architectures, ParaLIF’s accuracy matches or exceeds the highest performance reported in the literature on various neuromorphic datasets. These findings highlight ParaLIF as a promising approach for the development of rapid, accurate and energy-efficient SNNs, particularly well-suited for handling massive datasets containing long sequences.

A design of magnetic tunnel junctions for the deployment of neuromorphic hardware for edge computing

The electrically readable complex dynamics of robust and scalable magnetic tunnel junctions (MTJs) offer promising opportunities for advancing neuromorphic computing. In this work, we present an MTJ design with a free layer and two polarizers capable of computing the sigmoidal activation function and its gradient at the device level. This design enables both feedforward and backpropagation computations within a single device, extending neuromorphic computing frameworks previously explored in the literature by introducing the ability to perform backpropagation directly in hardware. Our algorithm implementation reveals two key findings: (i) the small discrepancies between the MTJ-generated curves and the exact software-generated curves have a negligible impact on the performance of the backpropagation algorithm, (ii) the device implementation is highly robust to inter-device variation and noise, and (iii) the proposed method effectively supports transfer learning and knowledge distillation. To demonstrate this, we evaluated the performance of an edge computing network using weights from a software-trained model implemented with our MTJ design. The results show a minimal loss of accuracy of only 0.4% for the Fashion MNIST dataset and 1.7% for the CIFAR-100 dataset compared to the original software implementation. These results highlight the potential of our MTJ design for compact, hardware-based neural networks in edge computing applications, particularly for transfer learning.

Spiking Variational Policy Gradient for Brain Inspired Reinforcement Learning.

Recent studies in reinforcement learning have explored brain-inspired function approximators and learning algorithms to simulate brain intelligence and adapt to neuromorphic hardware. Among these approaches, reward-modulated spike-timing-dependent plasticity (R-STDP) is biologically plausible and energy-efficient, but suffers from a gap between its local learning rules and the global learning objectives, which limits its performance and applicability. In this paper, we design a recurrent winner-take-all network and propose the spiking variational policy gradient (SVPG), a new R-STDP learning method derived theoretically from the global policy gradient. Specifically, the policy inference is derived from an energy-based policy function using mean-field inference, and the policy optimization is based on a last-step approximation of the global policy gradient. These fill the gap between the local learning rules and the global target. In experiments including a challenging ViZDoom vision-based navigation task and two realistic robot control tasks, SVPG successfully solves all the tasks. In addition, SVPG exhibits better inherent robustness to various kinds of input, network parameters, and environmental perturbations than compared methods.

Overcoming Neuroanatomical Mapping and Computational Barriers in Human Brain Synaptic Architecture.

In this Matters Arising, we critically examine the data processing and computational challenges highlighted under the high-resolution, three-dimensional reconstruction of human cortical tissue by Shapson-Coe et al. While the study represents a technical milestone in connectomics, involving a 1.4-petabyte dataset derived from mapping a cubic millimeter of temporal cortex, the findings also reveal the substantial obstacles inherent in scaling such approaches to the entire human brain. Beyond the application of artificial intelligence (AI) for segmentation and synapse detection, the study underscores the immense complexity of data acquisition, cleaning, alignment, and visualization at this scale. This article contextualizes these challenges by comparing the computational and infrastructural requirements of the Shapson-Coe work to other large-scale neuroscience initiatives, such as the fruit fly brain atlas, and explores emerging technologies like quantum computing and neuromorphic hardware as potential solutions. Additionally, we discuss the ethical and logistical implications of managing zettabyte-scale datasets and emphasize the necessity of international collaboration to achieve the ambitious goal of mapping the human connectome. By critically addressing these challenges and potential solutions, this article aims to guide future advancements in the field of connectomics and their transformative applications in neuroscience, artificial intelligence, and medicine.

Leveraging stochasticity in memristive synapses for efficient and reliable neuromorphic systems

Neuromorphic computing is inspired by the human brain’s architecture to develop power-efficient and optimized neural networks. Different characteristics of synapses and neurons are emulated in the neuromorphic hardware or software model to mimic the behavior of the synaptic brain. Noise in the neuromorphic system is an important constraint to evaluate its overall performance. A reliable READ operation is essential to secure an acceptable performance from a neuromorphic system. In addition, a memristive synapse is very prone to stochastic behavior. Moreover, a dynamically reconfigurable READ operation is proposed for our 3T1R synapse to explore the effect of stochasticity on neuromorphic applications. Our proposed method allows for relevant applications in neuromorphic computing to utilize the stochastic behavior of the memristive 3T1R synapse. Performance evaluations show ~8.9x energy optimization with our proposed device. Optimization algorithm (EONS) is utilized for training and a hardware framework (RAVENS) is used for hardware simulation during training.

Antiferromagnetic Skyrmion based Energy-Efficient Leaky Integrate and Fire Neuron Device.

The development of energy-efficient neuromorphic hardware using spintronic devices based on antiferromagnetic (AFM) skyrmion motion on nanotracks has gained considerable interest. Owing to their properties such as robustness against external magnetic fields, negligible stray fields, and zero net topological charge, AFM skyrmions follow straight trajectories that prevent their annihilation at nanoscale racetrack edges. This makes the AFM skyrmions a more favorable candidate over the ferromagnetic (FM) skyrmions for future spintronic applications. This work proposes an AFM skyrmion-based neuron device exhibiting the leaky-integrate-fire (LIF) functionality by exploiting either a thermal gradient or alternatively a perpendicular magnetic anisotropy (PMA) gradient in the nanotrack for leaky behavior by moving the skyrmion in the hotter or lower PMA region, respectively to minimize the system energy. Furthermore, it is shown that the AFM skyrmion couples efficiently to the soft FM layer of a magnetic tunnel junction (MTJ) enabling efficient read-out of the skyrmion. The maximum change of 9.2% in tunnel magnetoresistance (TMR) is estimated while detecting the AFM skyrmion. Moreover, the proposed neuron device has an energy dissipation of 4.32 fJ per LIF operation thus, paving the path for developing energy-efficient devices in AFM spintronics for neuromorphic computing.

A burst-dependent algorithm for neuromorphic on-chip learning of spiking neural networks

Abstract The field of neuromorphic engineering addresses the high energy demands of neural networks through brain-inspired hardware for efficient neural network computing. For on-chip learning with spiking neural networks, neuromorphic hardware requires a local learning algorithm able to solve complex tasks. Approaches based on burst-dependent plasticity have been proposed to address this requirement, but their ability to learn complex tasks has remained unproven. Specifically, previous burst-dependent learning was demonstrated on a spiking version of the ‘exclusive or’ problem (XOR) using a network of thousands of neurons. Here, we extend burst-dependent learning, termed ‘Burstprop’, to address more complex tasks with hundreds of neurons. We evaluate Burstprop on a rate-encoded spiking version of the MNIST dataset, achieving low test classification errors, comparable to those obtained using backpropagation through time on the same architecture. Going further, we develop another burst-dependent algorithm based on the communication of two types of error-encoding events for the communication of positive and negative errors. We find that this new algorithm performs better on the image classification benchmark. We also tested our algorithms under various types of feedback connectivity, establishing that the capabilities of fixed random feedback connectivity is preserved in spiking neural networks. Lastly, we tested the robustness of the algorithm to weight discretization. Together, these results suggest that spiking Burstprop can scale to more complex learning tasks and is therefore likely to be considered for self-supervised algorithms while maintaining efficiency, potentially providing a viable method for learning with neuromorphic hardware.

Optimizing electrochemical and ferroelectric synaptic devices: from material selection to performance tuning

Abstract Neuromorphic hardware systems emulate the parallel neural networks of the human brain, and synaptic weight storage elements are crucial for enabling energy-efficient information processing. They must represent multiple data states and be able to be updated analogously. In order to realize highly controllable synaptic devices, replacing the high-k gate dielectric in conventional transistor structures with either solid-electrolytes that facilitate bulk ionic motion or ferroelectric oxide allows for steady adjustment of channel currents in response to gate-voltage signals. This approach, in turn, accelerates backpropagation algorithms used for training neural networks. Furthermore, because the channel current in electrochemical random-access memory (ECRAM) is influenced by the number of mobile ions (e.g. Li+, O2−, H+ or Cu+) passing through the electrolytes, these synaptic device candidates have demonstrated an excellent linear and symmetrical channel current response when updated using an identical pulse scheme. In the latter case, which is known as the ferroelectric field-effect transistor (FeFET), the number of electrons accumulated near the channel rapidly varies with the degree of the alignment of internal dipoles in thin doped ferroelectric HfO2. This leads to a multilevel state. Based on the working principles of these two promising candidates, enabling gate-controlled ion-transport primarily in electrolytes for ECRAM and understanding the relationship between polarization and the ferroelectric layer in FeFETs are crucial to improve their properties. Therefore, this paper aims to present our recent advances, highlighting the engineering approaches and experimental findings related to ECRAM and FeFET for three-terminal synaptic devices.

A NEUROMORPHIC ANALYSIS OF CLIMATE PATTERNS FOR COMPLEX ENVIRONMENTAL FRACTAL MODELING

This study investigates the use of neuromorphic computing, particularly spiking neural networks (SNNs) and advanced neuromorphic hardware, to model and forecast climate patterns. Our neuromorphic system achieves high prediction accuracy, maintaining a Mean Squared Error (MSE) as low as 0.08, even with increasing data volumes. The system operates with notable energy efficiency, consuming just 0.15 J per inference at higher data loads. This efficiency, coupled with a throughput of 800 inferences per second, underscores the system’s capability to handle large-scale data effectively. The neuromorphic approach addresses key challenges in scalability and energy consumption, presenting a robust solution for real-time climate data analysis. By continuously adapting to new data inputs, the system ensures accurate and timely predictions, essential for applications in environmental monitoring and decision-making. The integration of artificial intelligence algorithms with neuromorphic architectures not only reduces computational costs but also enhances the interpretability of complex climate dynamics. These findings highlight the transformative potential of brain-like computing in environmental modeling, offering a scalable, efficient, and adaptable tool for climate prediction and analysis.

Heterogeneous Regularization for Fast Rendering Using Deep Spike Neural Network

A Deep Spiking Neural Network (DSNN) with Heterogeneous Regularization learning technique is proposed to build a more biologically plausible approach that evaluates the amount of noise and finds a stopping criterion for fast realistic illumination. Our contribution is to introduce a model that improves the label propagation of DSNN and is more efficient on neuromorphic hardware than a corresponding Artificial Neural Network. More specifically, we develop a biological neural model with a heterogeneous regularization technique that works similarly to a human brain and can detect noise using deep spikes without relying on mathematical metrics to extract noise features. The objective function of the proposed DSNN consists of a supervised term and an unsupervised term. The supervised term enforces the matching term between the predicted labels and the known labels. The unsupervised term enforces the smoothness of the predicted labels of the entire data samples. By learning a DSNN with the proposed objective function, we are able to develop a more powerful learning algorithm. Experiments were conducted using scenes with Global Illumination and various image distortions. The proposed model was also compared with the human visual system and other state-of-the-art models. The results show better performance and advantages in terms of efficiency, an increasingly biologically plausible network, and ease of implementation in Neuromorphic Hardware.

Neuromorphic Hardware for Artificial Sensory Systems: A Review

Abstract Senses are crucial for an organism’s survival, and there have been numerous efforts to artificially replicate sensory perception to elicit desired responses to specific stimuli. Recent research is increasingly focused on developing artificial sensory nervous systems based on the unsupervised learning capabilities of artificial neural networks (ANNs) using unstructured data. However, future ANNs, which require precise sensing capabilities in increasingly complex environments, must be capable of processing a large number of signals in real time, ideally from continuous domains. This need for massive data processing is driving the evolution of hardware systems, leading to the development of devices specifically designed for artificial sensory systems (ASSs) at the hardware level. To address this challenge, sensor devices need to not only detect target substances but also enable computational functions by utilizing their inherent material properties. Research in neuromorphic sensors is advancing towards integration with next-generation processing systems based on ANNs, effectively addressing the complex scenarios we aim to identify. This review offers perspectives on human-like sensor computing to address these challenges. It examines the progress in implementing five representative senses at the device level, explores methods for integrating them into systems for ASS, and provides a comprehensive overview of potential applications. In particular, we emphasize approaches to cognitively utilize the discussed devices as artificial sensory neurons and synapses, enabling responses to specific inputs. We aim to offer perspectives for the development of artificial sensory nerve systems in the future.

Distributed representations enable robust multi-timescale symbolic computation in neuromorphic hardware

Distributed representations enable robust multi-timescale symbolic computation in neuromorphic hardware

Critical band-to-band-tunnelling based optoelectronic memory

Neuromorphic vision hardware, embedded with multiple functions, has recently emerged as a potent platform for machine vision. To realize memory in sensor functions, reconfigurable and non-volatile manipulation of photocarriers is highly desirable. However, previous technologies bear mechanism challenges, such as the ambiguous optoelectronic memory mechanism and high potential barrier, resulting in a limited response speed and a high operating voltage. Here, for the first time, we propose a critical band-to-band tunnelling (BTBT) based device that combines sensing, integration and memory functions. The nearly infinitesimal barrier facilitates the tunnelling process, resulting in a broadband application range (940 nm). Furthermore, the observation of dual negative differential resistance (NDR) points confirms that the critical BTBT of photocarriers contributes to the sub-microsecond photomemory speed. Since the photomemory speed, with no motion blur, is important for motion detection, the critical BTBT memory is expected to enable moving target tracking and recognition, underscoring its superiority in intelligent perception.

Scalable network emulation on analog neuromorphic hardware.

We present a novel software feature for the BrainScaleS-2 accelerated neuromorphic platform that facilitates the partitioned emulation of large-scale spiking neural networks. This approach is well suited for deep spiking neural networks and allows for sequential model emulation on undersized neuromorphic resources if the largest recurrent subnetwork and the required neuron fan-in fit on the substrate. We demonstrate the training of two deep spiking neural network models-using the MNIST and EuroSAT datasets-that exceed the physical size constraints of a single-chip BrainScaleS-2 system. The ability to emulate and train networks larger than the substrate provides a pathway for accurate performance evaluation in planned or scaled systems, ultimately advancing the development and understanding of large-scale models and neuromorphic computing architectures.

Bioelectronics with Topological Crosslinked Networks for Tactile Perception

AbstractBioelectronics, which integrate biological systems with electronic components, have attracted significant attention in developing biomimetic materials and advanced hardware architectures to enable novel information‐processing systems, sensors, and actuators. However, the rigidity of conjugated molecular systems and the lack of reconfigurability in static crosslinked structures pose significant challenges for flexible sensing applications. Topological crosslinked networks (TCNs) featuring dynamic molecular interactions offer enhanced molecular flexibility and environmentally induced reconfigurability, decoupling the competition between performances. Here, recent advances are summarized in assembly methods of bioelectronics with different TCNs and elaborate ion/electron‐transport mechanisms from the perspective of molecular interactions. Decoupling effects can be achieved by comparing distinct TCNs and their respective properties, and an outlook is provided on a new range of neuromorphic hardware with biocompatibility, self‐healing, self‐powered, and multimodal‐sensing capabilities. The development of TCN‐based bioelectronics can significantly impact the fields of artificial neuromorphic perception devices, networks, and systems.

Ion Intercalation‐Mediated MoS2 Conductance Switching for Highly Energy‐Efficient Memristor Synapse

AbstractEmerging memristor synapses with ion dynamics have the potential to process spatiotemporal information and can accelerate the development of energy‐efficient neuromorphic computing. However, conventional ion‐migration‐type memristors suffer from low switching speed and uncontrollable conductance modulation, hindering energy‐efficient neuromorphic hardware implementation. Here, ion intercalation‐mediated conductance switching in MoS2 is introduced for a highly energy‐efficient memristor synapse (HEMS) to accurately emulate the bio‐synaptic function. Li‐ion intercalation into the few‐layer MoS2 can induce structural evolution, thereby achieving high‐speed and controllable conductance modulation in HEMS. Consequently, the HEMS exhibits highly energy efficiency with a fast switching speed of 500 ns and low energy consumption of 2.85 fJ per synaptic event. The stable bidirectional modulation of synaptic plasticity by consecutive voltage pulses of 5000 times can be achieved in the HEMS. Besides, the HEMS is endowed with logic functions and can process multiple sets of inputs in parallel for information integration. This work offers an alternative strategy for fast‐speed conductance modulation via ion intercalation to develop energy‐efficient memristors in future neuromorphic computing.

Rapid learning with phase-change memory-based in-memory computing through learning-to-learn

There is a growing demand for low-power, autonomously learning artificial intelligence (AI) systems that can be applied at the edge and rapidly adapt to the specific situation at deployment site. However, current AI models struggle in such scenarios, often requiring extensive fine-tuning, computational resources, and data. In contrast, humans can effortlessly adjust to new tasks by transferring knowledge from related ones. The concept of learning-to-learn (L2L) mimics this process and enables AI models to rapidly adapt with only little computational effort and data. In-memory computing neuromorphic hardware (NMHW) is inspired by the brain’s operating principles and mimics its physical co-location of memory and compute. In this work, we pair L2L with in-memory computing NMHW based on phase-change memory devices to build efficient AI models that can rapidly adapt to new tasks. We demonstrate the versatility of our approach in two scenarios: a convolutional neural network performing image classification and a biologically-inspired spiking neural network generating motor commands for a real robotic arm. Both models rapidly learn with few parameter updates. Deployed on the NMHW, they perform on-par with their software equivalents. Moreover, meta-training of these models can be performed in software with high-precision, alleviating the need for accurate hardware models.