Physical electronics for brain-inspired computing
类脑计算技术作为一种脑启发的新型计算技术, 具有存算一体、事件驱动、模拟并行等特征, 为智能化时代开发高效的计算硬件提供了技术参考, 有望解决当前人工智能硬件在能耗和算力方面的 “不可持续发展”问题. 硬件模拟神经元和突触功能是发展类脑计算技术的核心, 而支持这一切实现的基础是器件以及器件中的物理电子学. 根据类脑单元实现的物理基础, 当前类脑芯片主要可以分为数字 CMOS型、数模混合 CMOS型以及新原理器件型三大类. IBM的 TrueNorth、Intel的Loihi、清华大学的 Tianjic以及浙江大学的 Darwin等都是数字 CMOS型类脑芯片的典型代表, 旨在以逻辑门电路仿真实现生物单元的行为. 数模混合型的基本思想是利用亚阈值模拟电路模拟生物神经单元的特性, 最早由 Carver Mead提出, 其成功案例有苏黎世的 ROLLs、斯坦福的 Neurogrid等. 以上两种类型的类脑芯片虽然实现方式上有所不同, 但共同之处在于都是利用了硅基晶体管的物理特性. 此外, 以忆阻器为代表的新原理器件为构建非硅基类脑芯片提供了新的物理基础. 它们在工作过程中引入了离子动力学特性, 从结构和工作机制上与生物单元都具有很高的相似性, 近年来受到国内外产业界和学术界的广泛关注. 鉴于硅基工艺比较成熟, 当前硅基物理特性是类脑芯片实现的主要基础. 忆阻器等新原理器件的类脑计算技术尚处于前沿探索和开拓阶段, 还需要更成熟的制备技术、更完善的系统框架和电路设计以及更高效的算法等.
为促进本领域国内同行交流, 应《物理学报》编辑部邀请, 我们邀请部分活跃在利用物理器件实现类脑计算领域第一线的中青年科学家, 组织出版了本专题. 探讨不同物理机制的器件实现、算法优化、架构设计以及应用. 器件层面上, 刘琦老师报道了一种柔性忆阻器基神经元器件及电路, 徐文涛老师报道了一种基于 Na2/3Ni1/3Mn2/3O2的离子迁移型人造突触, 缪向水老师综述了主流忆阻器的器件结构、物理机制并比较分析了它们的性能特性, 万昌锦老师和万青老师介绍了多类柔性神经形态晶体管的研究进展以及在仿生感知领域中的应用; 算法层面上, 曾中明老师和袁喆老师利用磁性隧道结可调控的随机动力学实现了群体编码, 尚大山老师提出了一种基于存内计算范式的储池计算硬件实现方法; 系统层面上, 刘洋老师报道了一种基于忆阻器的脉冲神经网络硬件加速器架构设计方法, 郭新老师提出了一种仿生生物感官的感存算一体化系统, 康晋锋老师综述了用于实现存内计算的非挥发存储器件及其性能特征; 感知应用层面上, 王中强老师从器件物理角度、周菲迟老师和柴扬老师从应用角度分别综述了面向感存算功能一体化的光电忆阻器研究进展, 诸葛飞老师综述了光电神经形态器件及其应用的最新研究进展, 韩素婷老师综述了应用于感存算一体化系统忆阻器的研究方向和研究进展.
本专题从不同角度描述了面向类脑计算的器件及物理特性的进展, 反映了此领域的一些现状,希望对读者了解此前沿课题有所帮助, 可以吸引更多学者尤其是年轻学者的关注和加入, 为我国在本领域的蓬勃发展增添新生力量.
2022, 71 (14): 140501.
doi: 10.7498/aps.71.20220666
Abstract +
With the rapid development of deep learning, the current rapid update and iteration of intelligent algorithms put forward high requirements for hardware computing power. Limited by the exhaustion of Moore’s law and the von Neumann bottleneck, the traditional CMOS integration cannot meet the urgent needs of hardware computing power improvement. The utilization of new device memristors to construct a neuromorphic computing system can realize the integration of storage and computing, and has the characteristics of extremely high parallelism and ultra-low power consumption. In this work, the device structure and physical mechanism of mainstream memristors are reviewed in bottom-to-top order firstly, and their performance characteristics are compared and analyzed. Then, the recent research progress of memristors to realize artificial neurons and artificial synapses is introduced, including the simulation of specific circuit forms and neuromorphic functions. Secondly, in this work, the structural forms of passive and active memristive arrays and their applications in neuromorphic computing, including neural network-based handwritten digits and face recognition, are reviewed. Lastly, the current challenges of memristive brain-like computing from the bottom to the top, are summarized and the future development of this field is also prospected.
2022, 71 (14): 147301.
doi: 10.7498/aps.71.20220308
Abstract +
Biological perception system has the unique advantages of high parallelism, high error tolerance, self-adaptation and low power consumption. Using neuromorphic devices to emulate biological perceptual system can effectively promote the development of brain-computer interfaces, intelligent perception, biological prosthesis and so on. Compared with other neuromorphic devices, multi-terminal neuromorphic transistors can not only realize signal transmission and training learning at the same time, but also carry out nonlinear spatio-temporal integration and collaborative regulation of multi-channel signals. However, the traditional rigid neuromorphic transistor is difficult to achieve bending deformation and close fit with the human body, which limits the application range of neuromorphic devices. Therefore, the research of flexible neuromorphic transistor with good bending characteristics has become the focus of recent research. Firstly, this review introduces the research progress of many kinds of flexible neuromorphic transistors, including device structure, working principle and basic functions. In addition, the application of the flexible neuromorphic transistor in the field of bionic perception is also introduced. Finally, this review also gives a summary and simple prospect of the above research fields.
2022, 71 (14): 148504.
doi: 10.7498/aps.71.20220463
Abstract +
With the increasing demands for processing images and videos at edge terminals, complementary metal oxide semiconductor (CMOS) hardware systems based on conventional Von Neumann architectures are facing challenges in terms of energy consumption, speed, and footprint. Neuromorphic devices, including resistive random access memory with integrated storage-computation characteristic and optoelectronic resistive random access memory with highly integrated in-sensor computing characteristic, show great potential applications in image processing due to their high similarity to biological neural systems and advantages of high energy efficiency, high integration level, and wide bandwidth. These devices can be used not only to accelerate large numbers of computational tasks in conventional image preprocessing and higher-level image processing algorithms, but also to implement highly efficient biomimetic image processing algorithms. In this paper, we first introduce the state-of-the-art neuromorphic resistive random access memory and optoelectronic neuromorphic resistive random access memory, then review the hardware implementation of and challenges to image processing based on these devices, and finally provide perspectives of their future developments.
2022, 71 (14): 148505.
doi: 10.7498/aps.71.20220111
Abstract +
Conventional computers based on the von Neumann architecture are inefficient in parallel computing and self-adaptive learning, and therefore cannot meet the rapid development of information technology that needs efficient and high-speed computing. Owing to the unique advantages such as high parallelism and ultralow power consumption, bioinspired neuromorphic computing can have the capability of breaking through the bottlenecks of conventional computers and is now considered as an ideal option to realize the next-generation artificial intelligence. As the hardware carriers that allow the implementing of neuromorphic computing, neuromorphic devices are very critical in building neuromorphic chips. Meanwhile, the development of human visual systems and optogenetics also provides a new insight into how to study neuromorphic devices. The emerging optoelectronic neuromorphic devices feature the unique advantages of photonics and electronics, showing great potential in the neuromorphic computing field and attracting more and more attention of the scientists. In view of these, the main purpose of this review is to disclose the recent research advances in optoelectronic neuromorphic devices and the prospects of their practical applications. We first review the artificial optoelectronic synapses and neurons, including device structural features, working mechanisms, and neuromorphic simulation functions. Then, we introduce the applications of optoelectronic neuromorphic devices particularly suitable for the fields including artificial vision systems, artificial perception systems, and neuromorphic computing. Finally, we summarize the challenges to the optoelectronic neuromorphic devices, which we are facing now, and present some perspectives about their development directions in the future.
2022, 71 (14): 148502.
doi: 10.7498/aps.71.20220226
Abstract +
To develop future interactive artificial intelligence system, the construction of high-performance human perception system and processing system is vital. In a traditional perceptual and processing system, sensors, memory and processing units are physically separated because of their different functions and manufacture conditions, which results in frequent shuttling and format transformation of data resulting in long time delay and high energy consumption. Inspired by biological sensory nervous system, one has proposed the concept of in-sensor computing system in which the basic unit integrates sensor, storage and computing functions in the same place. In-sensor computing technology can provide a reliable technical scheme for the area of sensory processing. Artificial memristive synapse capable of sensing light, pressure, chemical substances, etc. is one type of ideal device for the application of in-sensor computing system. In this paper, at the device level, recent progress of sensory memristive synapses applied to in-sensor computing systems are reviewed, including visual, olfactory, auditory, tactile and multimode sensation. This review points out the challenge and prospect from the aspects of device, fabrication, integrated circuit system architecture and algorithms, aiming to provide possible research direction for future development of in-sensor computing system.
2022, 71 (14): 148501.
doi: 10.7498/aps.71.20220303
Abstract +
Neuromorphic electronics has received considerable attention recent years, and its basic functional units are synaptic electronic devices. A two-terminal artificial synapse with sandwiched structure emulates plasticity of the biological synapses under the action of nerve-like electrical impulse signals. In this paper, P3 phase Na2/3Ni1/3Mn2/3O2 multi-element metal oxides with layered structure are synthesized by sol-gel process. Owing to the fact that Na+ is easy to embed/eject into its crystal structure, an ion-migrating artificial synapse based on Na2/3Ni1/3Mn2/3O2 is designed and fabricated. The device emulates important synaptic plasticity, such as excitatory postsynaptic current, paired-pulse facilitation, spike-number dependent plasticity, spike-frequency dependent plasticity, spike-voltage amplitude dependent plasticity and spike-duration dependent plasticity. The device realizes the identification and response to Morse code commands.
2022, 71 (14): 148503.
doi: 10.7498/aps.71.20212323
Abstract +
Inspired by the working modes of the human brain, the spiking neuron plays an important role as the basic computing unit of artificial perception systems and neuromorphic computing systems. However, the neuron circuit based on complementary metal-oxide-semiconductor technology has a complex structure, high power consumption, and limited flexibility. These features are not conducive to the large-scale integration and the application of flexible sensing systems compatible with the human body. The flexible memristor prepared in this work shows stable threshold switching characteristics and excellent mechanical bending characteristics with bending radius up to 1.5 mm and bending times up to 104. The compact neuron circuit based on this device shows the key features of the neuron, such as threshold-driven spiking, all-or-nothing, refractory period, and strength-modulated frequency response. The frequency-input voltage relationship of the neuron shows the similarity of the rectified linear unit, which can be used to simulate the function of rectified linear unit in spiking neural networks. In addition, based on the electron transport mechanism, a core-shell model is introduced to analyze the working mechanism of the flexible memristor and explain the output characteristics of the neuron. In this model, the shell region consisting of Nb2O5–x is subjected to ohmic conduction, while the core region consisting of NbO2 is dominated by Poole-Frenkel conduction. These two mechanisms, combined with Newton’s law of cooling, dominate the threshold switching behavior of flexible memristor device. Furthermore, the threshold switching characteristic of the memristor is simulated, verifying the rationality of the working mechanism of the flexible memristor. Considering the fact that the threshold voltage decreases with temperature increasing, a correction term is added to the temperature of the shell region. Subsequently, the output characteristics of the neuron regulated by the input voltage are simulated. The simulation results show that the frequency increases but the threshold voltage decreases with the input voltage increasing, which is consistent with the experimental result. The introduction of the correction term confirms the influence of the thermal accumulation effect of the flexible substrate on neuron output characteristics. Finally, we build a spiking neural network based on memristive spiking neurons to implement handwriting recognition, achieving a 95.6% recognition rate, which is comparable to the ideal result of the artificial neural network (96%). This result shows the potential application of the memristive spiking neurons in neuromorphic computing. In this paper, the study of flexible neurons can guide the design of neuromorphic sensing and computing systems.
2022, 71 (14): 148507.
doi: 10.7498/aps.71.20220397
Abstract +
By integrating the storage and computing functions on the fundamental elements, computing in-memory (CIM) technology is widely considered as a novel computational paradigm that can break the bottleneck of Von Neumann architecture. Nonvolatile memory device is an appropriate hardware implementation approach of CIM, which possess significantly advantages, such as excellent scalability, low consumption, and versatility. In this paper, first we introduce the basic concept of CIM, including the technical background and technical characteristics. Then, we review the traditional and novel nonvolatile memory devices, flash and resistive random access memory (RRAM), used in non-volatile based computing in-memory (nvCIM) system. After that, we explain the operation modes of nvCIM: in-memory analog computing and in-memory digital computing. In addition, the applications of nvCIM are also discussed, including deep learning accelerator, neuromorphic computing, and stateful logic. Finally, we summarize the current research advances in nvCIM and provide an outlook on possible research directions in the future.
2022, 71 (14): 148702.
doi: 10.7498/aps.71.20220281
Abstract +
2022, 71 (14): 148401.
doi: 10.7498/aps.71.20220098
Abstract +
Spiking neural network (SNN) as the third-generation artificial neural network, has higher computational efficiency, lower resource overhead and higher biological rationality. It shows greater potential applications in audio and image processing. With the traditional method, the adder is used to add the membrane potential, which has low efficiency, high resource overhead and low level of integration. In this work, we propose a spiking neural network inference accelerator with higher integration and computational efficiency. Resistive random access memory (RRAM or memristor) is an emerging storage technology, in which resistance varies with voltage. It can be used to build a crossbar architecture to simulate matrix computing, and it has been widely used in processing in memory (PIM), neural network computing, and other fields. In this work, we design a weight storage matrix and peripheral circuit to simulate the leaky integrate and fire (LIF) neuron based on the memristor array. And we propose an SNN hardware inference accelerator, which integrates 24k neurons and 192M synapses with 0.75k memristor. We deploy a three-layer fully connected network on the accelerator and use it to execute the inference task of the MNIST dataset. The result shows that the accelerator can achieve 148.2 frames/s and 96.4% accuracy at a frequency of 50 MHz.
2022, 71 (14): 148701.
doi: 10.7498/aps.71.20220350
Abstract +
Neuromorphic computing system, inspired by human brain, has the capability of breaking through the bottlenecks of conventional von Neumann architecture, which can improve the energy efficiency of data processing. Novel neuromorphic electronic components are the hardware foundation of efficient neuromorphic computation. Optoelectronic memristive device integrates the functions of sensing, memorizing and computing and is considered as a promising hardware candidate for neuromorphic vision. Herein, the recent research progress of optoelectronic memristive device for in-sensor computing are reviewed, including optoelectronic materials and mechanism, optoelectronic memristive device/characteristics as well as functionality and application of in-sensor computing. We first review the optoelectronic materials and corresponding memristive mechanism, including photon-ion coupling and photon-electron coupling type. Then optoelelctronic and all-optical modulated memristive device are introduced according to the modulation mode. Moreover, we exhibit the applications of optoelectronic device in cognitive function simulation, optoelectronic logic operation, neuromorphic vision, object tracking, etc. Finally, we summarize the advantages/challenges of optoelectronic memristor and prospect the future development.
2022, 71 (14): 148506.
doi: 10.7498/aps.71.20220252
Abstract +
Developing suitable algorithms that utilize the natural advantages of the corresponding devices is a key issue in the hardware research of brain-inspired computing. Population coding is one of the computational schemes in biological neural systems and it contains the mechanisms for noise reduction, short-term memory and implementation of complex nonlinear functions. Here we show the controllable stochastic dynamical behaviors for the technically mature spintronic device, magnetic tunnel junctions, which can be used as the basis of population coding. As an example, we construct a two-layer spiking neural network, in which groups of magnetic tunnel junctions are used to code input data. After unsupervised learning, this spiking neural network successfully classifies the iris data set. Numerical simulation demonstrates that the population coding is robust enough against the nonuniform dispersion in devices, which is inevitable in fabrication and integration of hardware devices.
2022, 71 (14): 140701.
doi: 10.7498/aps.71.20220082
Abstract +
As a kind of brain-inspired computing, reservoir computing (RC) has great potential applications in time sequence signal processing and chaotic dynamics system prediction due to its simple structure and few training parameters. Since in the RC randomly initialized network weights are used, it requires abundant data and calculation time for warm-up and parameter optimization. Recent research results show that an RC with linear activation nodes, combined with a feature vector, is mathematically equivalent to a nonlinear vector autoregression (NVAR) machine, which is named next-generation reservoir computing (NGRC). Although the NGRC can effectively alleviate the problems which traditional RC has, it still needs vast computing resources for multiplication operations. In the present work, a hardware implementation method of using computing-in memory paradigm for NGRC is proposed for the first time. We use memristor array to perform the matrix vector multiplication involved in the nonlinear vector autoregressive process for the improvement of the energy efficiency. The Lorenz63 time series prediction task is performed by simulation experiments with the memristor array, demonstrating the feasibility and robustness of this method, and the influence of the weight precision of the memristor devices on the prediction results is discussed. These results provide a promising way of implementing the hardware NGRC.
2022, 71 (21): 210702.
doi: 10.7498/aps.71.20220974
Abstract +
A neuromorphic chip is an emerging AI chip. The neuromorphic chip is based on non-Von Neumann architecture, and it simulates the structure and working principle of the human brain. Compared with non-Von Neumann architecture AI chips, the neuromorphic chips have significant improvement of efficiency and energy consumption advantages. The 3D-NAND flash memory has the merits of a mature process and ultra-high storage density, and recently it attracted many researchers’ attention. However, owing to the proprietary nature of the technology, there are few hardware implementations. This paper reviews the present research status of neuromorphic computing by using the 3D-NAND flash memory, introduces the forward propagation and backward propagation schemes, and proposes several improvements on the device, structure, and architecture of 3D NAND for neuromorphic computing.
2022, 71 (21): 218504.
doi: 10.7498/aps.71.20221424
Abstract +
In recent years, the development of artificial intelligence has increased the demand for computing and storage. However, the slowing down of Moore’s law and the separation between computing and storage units in traditional von Neumann architectures result in the increase of power consumption and time delays in the transport of abundant data, raising more and more challenges for integrated circuit and chip design. It is urgent for us to develop new computing paradigms to meet this challenge. The neuromorphic devices based on the in-memory computing architecture can overcome the traditional von Neumann architecture by Ohm’s law and Kirchhoff’s current law. By adjusting the resistance value of the memristor, the artificial neural network which can mimic the biological brain will be realized, and complex signal processing such as image recognition, pattern classification and decision determining can be carried out. In order to further reduce the size of device and realize the integration of sensing, memory and computing, two-dimensional materials can provide a potential solution due to their ultrathin thickness and rich physical effects. In this paper, we review the physical effects and memristive properties of neuromorphic devices based on two-dimensional materials, and describe the synaptic plasticity of neuromorphic devices based on leaky integrate and fire model and Hodgkin-Huxley model in detail, including long-term synaptic plasticity, short-term synaptic plasticity, spiking-time-dependent plasticity and spiking-rate-dependent plasticity. Moreover, the potential applications of two-dimensional materials based neuromorphic devices in the fields of vision, audition and tactile are introduced. Finally, we summarize the current issues on two-dimensional materials based neuromorphic computing and give the prospects for their future applications.
