受生物神经系统启发的新型神经网络——脉冲神经网络，其神经元通过脉冲信号事件传递信息，不同于传统神经网络的值传递方式。脉冲神经网络的事件驱动特性保证了其在特定硬件架构上的低功耗优势。另一方面，新型阻变存储器（忆阻器）近来被广泛作为人工突触用于神经形态计算硬件的研究。基于忆阻器的硬件架构由于其存算一体特性以及并行模拟计算的高效性，是作为脉冲神经网络硬件实现的理想架构。本论文聚焦于在忆阻器阵列上实现脉冲神经网络应用，提出了一种创新的脉冲神经网络无监督训练算法（称为Greedy-STDP算法），通过时域脉冲信号的稀疏化以及脉冲神经网络的输入状态划分等方法，使得脉冲神经网络可以克服忆阻器硬件不可避免的器件非理想因素进行无监督训练。本论文通过手写体数字图像识别任务对提出的算法进行了仿真验证，结果表明Greedy-STDP算法大幅降低了脉冲神经网络对忆阻器器件中间状态数、一致性等特性的要求，同时Greedy-STDP算法也表现出了对忆阻器写噪声的容忍性，使得基于当前忆阻器器件实现脉冲神经网络在线训练成为可能。
The next-generation neural network, Spiking Neural Network (SNN), which is inspired by biological neural systems, conveys information by spike events between neurons, while neurons in conventional artificial neural networks communicate with each other by passing values. The event-driven property of SNNs offers great potential for implementing energy-efficient applications on specific hardwares. On the other hand, the emerging Resistive Random Access Memory (RRAM), also known as memristors, have become popular artificial synapse candidates for neuromorphic computing hardwares. The computing-in-memory and highly-parallel-analog-computing property of memristors, makes RRAM an appropriate hardware implementation for SNNs. This article focuses on RRAM-based hardware implementations for SNN applications, and proposes a novel unsupervised training method (called Greedy-STDP algorithm) for SNNs by sparsifing spike event in the temporal domain and dividing input phases, which makes unsupervised SNN training tolerant with inevitable non-idealities of RRAM devices. We also conduct handwritten digit recognition simulations for the proposed training algorithm. The results have shown that Greedy-STDP algorithm could greatly mitigate the requirement for the intermediate level number and consistency of RRAM devices, and the Greedy-STDP training process also shows immunity to the write variations of RRAM devices. The proposed Greedy-STDP algorithm makes unsupervised SNN training more compatible with non-idealities of current RRAM devices, promising high feasibility of RRAM-based neuromorphic hardwares for online training.
