研究背景
人脑视觉皮层中神经网络的发展支持了人类视觉的记忆性。通过识别和处理实时视觉信息进行自主控制的人工视觉系统的进步，鼓励了这种生态视觉感知。与生物视觉系统类似人工视觉系统是由收集视觉数据的光电探测器、存储成像数据的记忆单元和处理图像、进行神经形态计算和识别任何物体的处理器组成。为了部署有效的类脑视觉系统，必须实现单一的成像单元，并结合内置存储器和信号处理能力。由于其天生的光学敏感性和记忆特性，光电记忆和突触装置已被探索为人工视觉系统的可行装置。这些设备在最近的创造性突破中显示了为机器视觉进行图像预处理和神经形态计算的能力。传统的CMOS 人工视觉系统由于其功能部件包括感觉终端、计算和存储单元的分离，存在着高功耗、数据访问延迟和硬件终止等重大挑战，这对其适应未来神经形态系统的能力构成重大障碍。在人类视觉结构的激励下，光学功能的神经形态记忆装置已经被开发出来，用于神经形态视觉传感器。这些器件在神经形态视觉传感器中表现出极大的关注，因为它们在同一个光学忆阻器中结合了光学传感、记忆和处理功能。光子/光学忆阻器的这些优良特性降低了计算延迟和功耗。
研究成果
光电子记忆突触装置以其在可见光脉冲和电信号下工作而闻名，在神经形态计算系统和人工视觉信息处理方面有很好的潜力。在此，阿卜杜拉国王科技大学Nazek El-Atab教授团队提出了一种基于可溶液加工的黑磷/氧化亚铁双层的灵活的线后兼容光电记忆器，具有良好的突触功能，面向仿生视网膜。该装置显示了高度稳定的突触特征，如长期电位(LTP)和长期抑制(LTD)，重复1000 次每次有 400个传导脉冲。该装置在长期记忆(LTM)/短期记忆(STM)方面呈现出先进的突触特征，以及在可见光诱导下的学习-遗忘-再学习。这些先进的突触功能可以提高神经形态应用的信息处理能力。有趣的是，通过调整光的强度和照明时间，STM 可以转换为 LTM。利用该设备的光诱导特性，开发了一个 6x6 的突触阵列，以展示在人工视觉感知中的可能用途。此外，该装置使用硅反蚀工艺进行了弯曲。所得的柔性器件在弯曲到 1厘米半径时表现出稳定的突触特征。单个记忆单元的这些多功能特性使其非常适用于光电记忆存储、神经形态计算和人工视觉感知应用。相关研究以“Flexible Solution Processable Black Phosphorus Based Optoelectronic Memristive Synapse for Neuromorphic Computing and Artificial Visual Perception Applications”为题发表在Advanced Materials期刊上。
图文导读
Figure. 1 a). The schematic structure of bionic human vision system and photonic memristive array for neuromorphic vision system. b) and c) The cross-section TEM images of the devices with the scale bar of 200 nm and 100 nm, respectively. d) EDS elemental line profile to validate the various elements in the device. e) EDS mapping of various elements Cu, P, Hf, O, and Pt presented in the Cu/HfOx/BP/Pt memristive device. f) XRD spectrum of BP. g) XPS analysis of BP.
Figure. 2 a). I-V characteristics of the HfOx based device for 1st, 10th, 50th, 80th, and 100th cycles. b) DC endurance of the HfOx device. c) SET/RESET voltages distribution with cycles for HfOx device (inset shows the statistical threshold voltage distribution). d) I-V curves of the HfOx/BP based device for 1st, 10th, 50th, 80th, and 100th cycles. e) DC endurance of the HfOx/BP device. f) SET/RESET voltages distribution with cycles for HfOx/BP device (inset shows the statistical threshold voltage distribution). g) multilevel states of the HfOx/BP device with reset stop voltages from -1.2 V to -1.5 V. h) multilevel retention states of the HfOx/BP device. i) The AC endurance of the HfOx/BP memristive device with the speed of 100 ns for both SET and RESET.
Fig. 3. The top-view optical image of the Cu/HfOx/BP/Pt devices.
Fig. 4 (a). Schematic structure of the fresh device. (b)-(d) Schematic diagrams of random growth and breach of the conductive filament in the Cu/HfOx/Pt device.
Figure 5. a). Potentiation and depression of the HfOx/BP device by 400 SET pulses and 400 RESET pulses. b) Potentiation and depression of HfOx/BP device by SET/RESET training pulses showing the analog ON/OFF conductance ratio. The pulse train consists of consecutive SET pulses (0.9 V, 10 µs) followed by consecutive RESET pulses (−0.8 V, 10 μs). The pulse trains include 100/150/200/300/400 SET pulses and 100/150/200/300/400 RESET pulses, respectively. c) Linear potentiation and depression. The pulse train consists of 100 set pulses followed by 100 reset pulses. d) The PPF index of the device. Inset: schematic illustration of the measurement of paired-pulse facilitation. e, f) Repeatability of LTP and LTD performance of device for total number of 1000 epochs, where each cycle has 800 conductance states (400 for potentiation and 400 for depression) corresponding to applied AC pulses. g) The ANN network structure consists of input, hidden and output layers. h) The recognition accuracy increases with the training epochs. i) Confusion matrix with different training states.
Fig. 6 (a). Schematic design of the biological synapse and artificial optoelectronic memristive synapse. b) The light[1]induced photo synaptic current using the single blue light and electrical erase using the voltage pulse. c) The PPF index variation with the time interval (Δt) of photonic pulse pairs. d) Photo synaptic current response under the light for 5 swith numerous intensities of light, followed by photo synaptic current decay when light is turned off. e) Photo synaptic current response under the light intensity of 68 mW/cm2 with different time duration (2 s, 3 s, 4s, 5 s), followed by photocurrent decay when light is turned off. f). The multi-storage psychological paradigm of human memory system, which consists of three kinds of memories, that are, sensory memory (SM), short-term memory (STM) and long-term memory (LTM). This is also a learning-forgetting[1]relearning procedure. g) Memory strength model in the device, which is inspired by the multi-storage model. h) The learning-forgetting-relearning procedure for seven cycles.
Figure 7. a) The Schematic of giving light pulse to sample b) Image mapping of the 6x6 array with five different light intensities c) The conductance mapping to the 6x6 array in both potentiation and depression state with corresponding time at 0, 3 and 6 s.
Figure 8. a) The photographic image of the flexible device during the electrical measurements. b) the LRS/HRS current of the flexible device with bending radius. c) The LRS/HRS states of the device with the bending radius of R=1 cm. d) LTP and LTD Electrical SET/Optical RESET endurance. d) Repeatability of LTP and LTD characteristics of flexible device with total number of 25 epochs.
总结与展望
HfOx/BP的光电突触行为可以作为神经形计算系统的硬件实现，具有低功耗和处理速度的优势。该装置显示了高度稳定的突触特征，如LTP 和LTD，重复 1000 次，每次400 个电导脉冲。在可见光诱导下，该装置在 LTM、STM、STP 和学习-遗忘-再学习方面显示出先进的突触特征。为了突出光学学习-遗忘-再学习过程的优势，最终构建了一个由 HfOx/BP 组成的 6x6 光电突触阵列，并用于模拟人类的视觉感知和视觉记忆功能。光学传感和突触学习行为的综合功能使其成为未来神经形态计算系统以及人工视觉和可穿戴应用的光电记忆存储的候选者。
文献链接
Flexible Solution Processable Black Phosphorus Based Optoelectronic Memristive Synapse for Neuromorphic Computing and Artificial Visual Perception Applications
https://doi.org/10.1002/adma.202300446
转自：“i学术i科研”微信公众号
如有侵权，请联系本站删除！
