The ever-increasing demand for training and inferring with larger machine-learning models requires more efficient hardware solutions due to limitations such as power dissipation and scalability. Optics is a promising contender for providing lower power computation, since light propagation through a nonabsorbing medium is a lossless operation. However, to carry out useful and efficient computations with light, generating and controlling nonlinearity optically is a necessity that is still elusive. Multimode fibers (MMFs) have been shown that they can provide nonlinear effects with microwatts of average power while maintaining parallelism and low loss. We propose an optical neural network architecture that performs nonlinear optical computation by controlling the propagation of ultrashort pulses in MMF by wavefront shaping. With a surrogate model, optimal sets of parameters are found to program this optical computer for different tasks with minimal utilization of an electronic computer. We show a remarkable decrease of 97% in the number of model parameters, which leads to an overall 99% digital operation reduction compared to an equivalently performing digital neural network. We further demonstrate that a fully optical implementation can also be performed with competitive accuracies.

Artificial synapses utilizing spike signals are essential elements of new generation brain-inspired computers. In this paper, we realize light-stimulated adaptive artificial synapse based on nanocrystalline zinc oxide film. The artificial synapse photoconductivity shows spike-type signal response, long and short-term memory (LTM and STM), STM-to-LTM transition and paired-pulse facilitation. It is also retaining the memory of previous exposures and demonstrates spike-frequency adaptation properties. A way to implement neurons with synaptic depression, tonic excitation, and delayed accelerating types of response under the influence of repetitive light signals is discussed. The developed artificial synapse is able to become a key element of neuromorphic chips and neuromorphic sensorics systems.

类脑神经形态计算通过电子或光子器件集成来模拟人脑结构和功能。人工突触是类脑系统中数量最多的计算单元。忆阻器可模拟突触功能, 并具有优异的尺寸缩放性和低能耗, 是实现人工突触的理想元器件。利用欧姆定律和基尔霍夫定律, 忆阻器交叉阵列可执行并行的原位乘累加运算, 从而大幅提升类脑系统处理模拟信号的速度。氧化物制备容易, 和CMOS工艺兼容性强, 是使用最广泛的忆阻器材料。本文梳理了氧化物忆阻器的研究进展, 分别讨论了电控、光电混合调控和全光控忆阻器, 主要聚焦阻变机理、器件结构和性能。电控忆阻器工作一般会产生微结构变化和焦耳热, 将严重影响器件稳定性, 改进器件结构和材料成分可有效改善器件性能。利用光信号调控忆阻器电导, 不仅能降低能耗, 而且可避免产生微结构变化和焦耳热, 从而有望解决稳定性难题。此外, 光控忆阻器能直接感受光刺激, 单器件即可实现感/存/算功能, 可用于研发新型视觉传感器。因此, 全光控忆阻器的实现为忆阻器的研究和应用打开了一扇新窗口。

研制具有生物神经元信息功能的柔性电子器件对于发展智能穿戴技术具有重要意义。传统阈值型忆阻器可模仿神经元信息整合功能, 但因缺乏本征柔韧性, 难以满足应用需求。本工作制备了一种基于本征可拉伸阈值型忆阻器的柔性人工神经元, 它由银纳米线-聚氨酯复合介质薄膜和液态金属电极构成。在外加电压下, 器件呈现良好的阈值电阻转变特性, 这归因于银纳米线间形成非连续银导电细丝的动态通断。该器件可模仿生物神经元的信息整合-发放及脉冲强度和脉冲间隔调制的尖峰放电功能。在20%拉伸应变下, 器件工作参数基本保持稳定, 性能未发生明显退化。本工作为发展可拉伸柔性人工神经元及下一代智能穿戴设备提供重要材料和技术参考。

忆阻器可以在单一器件上实现存储和计算功能, 成为打破冯·诺依曼瓶颈的核心电子元器件之一。它凭借独特的易失性/非易失性电阻特性, 可以很好地模拟大脑活动中的突触/神经元的功能。此外, 基于金属氧化物的忆阻器与传统的互补金属氧化物半导体(CMOS)工艺兼容, 受到了广泛关注。近年来, 研究提出了多种基于单介质层结构的金属氧化物忆阻器, 但仍然存在高低阻态不稳定、开关电压波动大和循环耐久性差等问题。在此基础上, 研究人员通过在金属氧化物忆阻器中引入双介质层成功优化了忆阻器的性能。本文首先详细介绍了氧化物双介质层忆阻器的优势, 阐述了氧化物双介质层忆阻器的阻变机理和设计思路, 并进一步介绍了氧化物双介质层忆阻器在神经形态计算中的应用。本文将为设计更高性能的氧化物双介质层忆阻器起到一定的启示作用。

With rapid advancement and deep integration of artificial intelligence and the internet-of-things, artificial intelligence of things has emerged as a promising technology changing people’s daily life. Massive growth of data generated from the devices challenges the AIoT systems from information collection, storage, processing and communication. In the review, we introduce volatile threshold switching memristors, which can be roughly classified into three types: metallic conductive filament-based TS devices, amorphous chalcogenide-based ovonic threshold switching devices, and metal-insulator transition based TS devices. They play important roles in high-density storage, energy efficient computing and hardware security for AIoT systems. Firstly, a brief introduction is exhibited to describe the categories (materials and characteristics) of volatile TS devices. And then, switching mechanisms of the three types of TS devices are discussed and systematically summarized. After that, attention is focused on the applications in 3D cross-point memory technology with high storage-density, efficient neuromorphic computing, hardware security (true random number generators and physical unclonable functions), and others (steep subthreshold slope transistor, logic devices,etc.). Finally, the major challenges and future outlook of volatile threshold switching memristors are presented.

二维过渡金属硫化合物是构建纳米电子器件的理想材料, 基于该材料体系开发用于信息存储和神经形态计算的忆阻器, 受到了学术界的广泛关注。受制于低成品率和低均一性问题, 二维过渡金属硫化合物忆阻器阵列鲜见报道。本研究采用化学气相沉积得到厘米级二维碲化钼薄膜, 并通过湿法转移和剥离工艺制备得到碲化钼忆阻器件。该碲化钼器件表现出优异的保持性(保持时间>500 s)、快速的阻变(SET时间~60 ns, RESET时间~280 ns)和较好的循环寿命(阻变2000圈后仍可正常工作)。该器件具有高成品率(96%)、低阻变循环间差异性(SET过程为6.6%, RESET过程为5.2%)和低器件间差异性(SET过程为19.9%, RESET过程为15.6%)。本工作成功制备出基于MoTe₂的3×3忆阻器阵列。在此基础上, 将研制的MoTe₂器件用于手写体识别, 实现了91.3%的识别率。最后, 通过对MoTe₂器件高低阻态的电子输运机制进行拟合分析, 揭示了该器件阻变源于类金属导电细丝的通断过程。本项工作表明大尺寸二维过渡金属硫化合物在未来神经形态计算中具有巨大的应用潜力。

脑科学与类脑研究是国际必争战略性前沿。人工智能与深度学习的飞速发展对算力提出了迫切需求。而传统的冯诺依曼架构，由于存算分离导致功耗墙和内存墙，摩尔定律也逐渐放缓。光神经拟态计算充分融合高速光通信、光互连、光集成、硅基光电子与神经拟态计算的特点，具有超高速、大带宽、多维度等优势，在高性能计算、人工智能领域有广阔的应用前景，是突破后摩尔时代传统微电子计算极限极具竞争力的方案。本文回顾了国内外主要研究团队在光神经元、光突触、光神经网络的理论、算法及器件方面的工作，并提出了展望。

结合时延储备池计算和垂直腔面发射激光器，基于现有的光纤光学平台，对以1 550 nm波段垂直腔面发射激光器为非线性节点的时延光储备池计算系统进行了实验研究。结果表明，在该实验系统中可以分别成功地实现单个Santa-Fe混沌时间序列预测任务以及单个非线性信道均衡任务。基于垂直腔面发射激光器在特定参数条件下能实现双模共存，进一步在该系统中垂直腔面发射激光器的两个偏振模式中同时注入外部信号，成功地完成了Santa-Fe混沌时间序列预测和非线性信道均衡任务的并行处理，但是整体性能要弱于单任务处理性能；除此之外，并行任务的性能随着外光注入强度的增加而得到改善，其中注入强度比率大的一方性能更好。