东北师范大学刘益春校长
在国家自然科学基金及国家重大科学研究计划的资助下，东北师范大学刘益春研究组利用InGaZnO材料，构造了具有自主学习和记忆能力的神经突触仿生器件，在单一无机器件中实现了多种生物突触功能。相关成果发表在国际学术期刊《先进功能材料》上，并被选为标题页文章进行了重点报道。
据介绍，神经突触是人类大脑学习和记忆的基本组成单元，突触仿生是实现神经形态计算的重要基础。突触可以看做是一种两端器件，其突触权重可对刺激信号作出动态响应，这一特点恰恰与忆阻器的概念相似——电阻的阻值可以随流经电量而发生改变。因此，利用忆阻器件实现对神经突触的仿生一直是相关领域的研究热点。
在东北师范大学教授刘益春的带领下，该研究组利用非晶态InGaZnO薄膜的电学性质可调节性及其对激励信号可作出动态反应等特点，设计并制备了由两层不同含氧量的InGaZnO薄层构成的忆阻器件；实现了对神经突触多种生物功能的模拟，涉及兴奋性突触后电流、非线性传输特性、长时程/短时程可塑性、刺激频率响应特性、STDP机制、经验式学习等多个方面，尤其是器件表现出的短时记忆行为与“学习—忘记—再学习”的经验式学习模式符合人类的认知规律。
同时，科研人员通过系统研究短时可塑性随温度的变化规律，揭示了该器件的运行机制为氧离子的迁移和扩散。
相关专家表示，该成果对促进更加精确地仿生神经突触进而实现人工神经网络打下了坚实的基础。
Synaptic Learning and Memory Functions Achieved Using Oxygen Ion Migration/Diffusion in an Amorphous InGaZnO Memristor
Zhong Qiang Wang, Hai Yang Xu, Xing Hua Li, Hao Yu, Yi Chun Liu, Xiao Juan Zhu
A single synaptic device with inherent learning and memory functions is demonstrated based on an amorphous InGaZnO (α-IGZO) memristor; several essential synaptic functions are simultaneously achieved in such a single device, including nonlinear transmission characteristics, spike-rate-dependent and spike-timing-dependent plasticity, long-term/short-term plasticity (LSP and STP) and “learning-experience” behavior. These characteristics bear striking resemblances to certain learning and memory functions of biological systems. Especially, a “learning-experience” function is obtained for the first time, which is thought to be related to the metastable local structures in α-IGZO. These functions are interrelated: frequent stimulation can cause an enhancement of LTP, both spike-rate-dependent and spike-timing-dependent plasticity is the same on this point; and, the STP-to-LTP transition can occur through repeated “stimulation” training. The physical mechanism of device operation, which does not strictly follow the memristor model, is attributed to oxygen ion migration/diffusion. A correlation between short-term memory and ion diffusion is established by studying the temperature dependence of the relaxation processes of STP and ion diffusion. The realization of important synaptic functions and the establishment of a dynamic model would promote more accurate modeling of the synapse for artificial neural network.
