三维忆阻递归神经网络的遗忘相关联想记忆

Associate memory and forgetting ability are two important traits for human memory, however compared with the associate memory, the forgetting ability is rarely concerned in the research of the traditional recurrent neural networks (RNN). This project is mainly studied the associate memory that affected by the forgetting effect of the 3-dimensional (3D) memristor in the RNN. The 3D memristor is preferred to be a more accurate and general model, the parameter conditions for memristors with unipolar, bipolar, reversible unipolar and bipolar, bipolar with forgetting effect are obtained respectively. The dynamical of the RNN based on the 3D memristor model, is related to both the input and the time. The stability conditions of the RNN are discussed by dividing the system into various subsystems in terms of the special properties of the 3D memristor. The time delay and the robust stability of the RNN are further studied. The neuromorphic circuit of the RNN based on the memristor and the neuron emulators can not only verify the analysis of the memristive RNN, but also provide a reference for the memristor based electronic brain.

联想记忆和遗忘特性是人类记忆的两个重要特点，传统递归神经网络主要研究联想记忆，而忽略了遗忘特性。本项目主要研究具有遗忘效应的三维忆阻器构建的递归神经网络的联想记忆特性。分析三维HP模型和三维肖特基模型表现出单极，双极，单双极可逆，以及具有遗忘效应的忆阻器的参数条件，并以实物忆阻器数据进行验证。以三维忆阻器构建的递归神经网络联想记忆不仅与输入相关，还与时间相关，结合三维忆阻器的阈值特性和时间相关特性，将时变输入相关递归神经网络划分为不同子系统进行稳定性分析，并进一步考虑时滞和扰动等的影响。结合三维忆阻器和神经元模拟器构建忆阻递归网络的神经形态电路，突破传统的电路结构，在同一电路上实现自联想和异联想记忆，并对理论分析结果加以验证，研究成果可直接应用于构建基于忆阻器的电子大脑。

类脑神经网络的实现需要结构简单，行为丰富，功耗低，便于集成和扩展的元件。突触和神经元是神经网络的两个基本部件，纳米级器件忆阻器作为天然的突触，自成多元动力系统，表现出15种行为特性。神经元也是独立的动力系统，表现出20种生物脉冲信号模式。忆阻器和神经元两个动力系统之间的信号耦合，使神经网络自发形成时间相关可塑性学习规则，因此不需外围电路，即可控制网络的连接结构和学习功能。通过对忆阻器、神经元和两者间的时间相关可塑性、对联想记忆神经网络性能的影响进行分析，为类脑神经网络的设计和实现提供参考。发表SCI文章8篇，EI会议论文1篇，获得国家授权发明专利1项。获得实用新型专利1项。其中获得本项目资助的SCI文章8篇，会议论文1篇。参加国际会议1此。邀请国内外专家学术交流3次，