搜索

x
中国物理学会期刊

可用于图像降噪预处理的MoTe2/WS2异质结光电突触晶体管

MoTe2/WS2 Heterojunction Optoelectronic Synapse Transistor for Image Denoising Preprocessing

PDF
导出引用
  • 面向后摩尔时代计算需求,开发兼具低功耗和多功能集成特性的神经形态计算硬件,对于突破传统冯·诺依曼架构限制、提升计算能效具有重要意义。在此,我们报道了一种基于MoTe2/WS2异质结的多功能集成光电突触,在单一器件中实现了从基础突触行为模拟到复杂视觉信息处理的功能集成。器件在光刺激下表现出丰富的突触可塑性行为,单次读出电路能耗低至3.29 pJ;通过背栅电压与光脉冲的协同调控,成功演示了光适应行为及可重构逻辑运算。利用器件由短期向长期可塑性的过渡特征实现了图像降噪功能,对加噪MNIST数据集的识别准确率从72.33%显著回升至93.27%。这些结果表明该器件可用于构建低功耗神经形态计算与多功能视觉系统。

     

    To meet the computational demands of the post-Moore era, neuromorphic hardware with low-power operation and multifunctional integration is highly desirable for overcoming the von Neumann bottleneck and enhancing computational efficiency. Here, we present a multifunctional optoelectronic synaptic transistor based on a MoTe2/WS2 heterojunction, integrating functionalities ranging from fundamental synaptic emulation to advanced visual information processing within a single device. First, under optical stimulation, the device exhibits fundamental synaptic functions such as paired-pulse facilitation, pulse-number-dependent plasticity, pulse-frequency-dependent plasticity, and experience-dependent learning, with an ultralow energy consumption of 3.29 pJ per readout event. Furthermore, the device successfully emulates the visual light adaptation behavior of the human eye. Under strong illumination, applying a negative gate voltage reduces the PSC, mimicking the contraction of the pupil under intense light to reduce incident light, thereby maintaining visual comfort and protecting the retina. Beyond this, by defining the optical pulse intensity and back-gate voltage as two independent logic inputs and the corresponding PSC as the logic output, the device enables reconfigurable logic operations that can be switched between OR and AND functions by tuning the source-drain bias. Finally, the transition from short-term plasticity to long-term plasticity of the device is exploited for image denoising. Following device-based denoising preprocessing, the recognition accuracy of the Gaussian-noise-corrupted MNIST dataset increases from an initial 72.33% to 93.27%, approaching the accuracy achieved with the raw dataset. These results highlight the potential of the device for low-power neuromorphic computing and multifunctional visual information processing.

     

    目录

    /

    返回文章
    返回