Hou Zhidong, Shen Jinrong, Zhong Yiming, Wu Dongping
School of Microelectronics, Fudan University, Shanghai 200433, China.
Micromachines (Basel). 2025 Apr 28;16(5):517. doi: 10.3390/mi16050517.
Brain-inspired neuromorphic computing is expected to overcome the von Neumann bottleneck by eliminating the memory wall between processing and memory units. Nevertheless, critical challenges persist in synaptic device implementation, particularly regarding nonlinear/asymmetric conductance modulation and multilevel conductance states, which substantially impede the realization of high-performance neuromorphic hardware. This study demonstrates a novel advancement in photonic-electronic modulated synaptic devices through the development of an amorphous indium-gallium-zinc oxide (a-IGZO) synaptic transistor. The device demonstrates biological synaptic functionalities, including excitatory/inhibitory post-synaptic currents (EPSCs/IPSCs) and spike-timing-dependent plasticity, while achieving excellent conductance modulation characteristics (nonlinearity of 0.0095/-0.0115 and asymmetric ratio of 0.247) and successfully implementing Pavlovian associative learning paradigms. Notably, systematic neural network simulations employing the experimental parameters reveal a 93.8% recognition accuracy on the MNIST handwritten digit dataset. The a-IGZO synaptic transistor with photonic-electronic co-modulation serves as a potential critical building block for constructing neuromorphic architectures with human-brain efficiency.
受大脑启发的神经形态计算有望通过消除处理单元和存储单元之间的存储墙来克服冯·诺依曼瓶颈。然而，在突触器件的实现方面仍然存在关键挑战，特别是在非线性/非对称电导调制和多级电导状态方面，这严重阻碍了高性能神经形态硬件的实现。本研究通过开发一种非晶铟镓锌氧化物（a-IGZO）突触晶体管，展示了光子-电子调制突触器件的一项新进展。该器件展示了生物突触功能，包括兴奋性/抑制性突触后电流（EPSCs/IPSCs）和脉冲时间依赖可塑性，同时实现了优异的电导调制特性（非线性为0.0095/-0.0115，不对称比为0.247），并成功实现了巴甫洛夫联想学习范式。值得注意的是，使用实验参数进行的系统神经网络模拟在MNIST手写数字数据集上显示出93.8%的识别准确率。具有光子-电子协同调制的a-IGZO突触晶体管是构建具有人脑效率的神经形态架构的潜在关键组件。
