面向类脑记忆的忆阻神经网络及其在联想记忆中的应用

In recent years, exploring the working mechanism of human brain and establishing artificial neural network with brain-inspired characteristics are one of the hottest research directions of artificial intelligence field. Making use of the current research results of brain science and neuroscience, and learning from the existing theory of long-short-term memory, this project will design dynamic synaptic adjustment rules and synaptic plasticity algorithms based on brain-inspired memory. We will also investigate neuronal activation functions and the neural network topological structure, and establish a neural network model of brain-inspired memory, so that the self-transformation from short-term memory network to long-term memory network will be realized. Designing synaptic circuits, neuron circuits and neural network circuits by using memristors as electronic synapses, and the hardware implementation of the neural network of the brain-inspired memory will come true. We will also explore the applications of the brain-inspired memory in associative memory. Long-term and short-term memory network models are designed respectively. Short-term memory network dynamically learns model information, while long-term memory network stores and recognizes model information. Through the division and cooperation of the two networks, the problems in traditional associative memory application such as small memory capacity and insufficient capacity of associative memory processing in mixed modes are solved. This study will set up a whole set of neural network design, implementation and application research system based on brain-inspired memory. The research results will play an important role in promoting neural networks, brain-inspired intelligence, memristor and associative memory. It is conducive to promote our country's research in the field of brain-inspired intelligence theory and technology.

探索人脑工作机制，建立具有类脑特性的神经网络是人工智能领域近年来的热门研究方向。利用当前脑科学与神经科学的研究成果，借鉴现有的人脑长短时记忆理论，设计面向类脑记忆的动态突触调整法则与突触可塑性算法，研究神经网络激活函数与拓扑结构，建立面向类脑记忆的神经网络模型，实现短时记忆网络到长时记忆网络的自主转换。以忆阻作为电子突触，设计基于忆阻的突触、神经元与神经网络等电路，硬件实现面向类脑记忆的神经网络。探索类脑记忆神经网络在联想记忆中的应用，设计长短时记忆网络模型，短时记忆网络动态学习模式信息，长时记忆网络存储与识别模式信息，通过两种网络的分工与合作，解决传统联想记忆应用中的记忆容量小、混合模式处理能力不足等问题。本研究将建立面向类脑记忆的神经网络设计、实现及应用研究体系，研究成果对神经网络、类脑智能、忆阻与联想记忆等研究都将产生重要的推动作用，有利于促进我国在类脑智能领域的理论研究与技术发展。

忆阻作为一种新的纳米级器件，本质上是一个阻值可调的存储器，特别适合实现人工神经网络中的突触权值，给神经网络中突触权值的调整带来新的方式。本项目围绕忆阻突触、神经元电路的设计、忆阻神经网络系统的实现、神经网络动力学行为分析与控制，以及基于忆阻的联想记忆的实现等开展了深入研究。.设计了电压和电流模式的忆阻突触电路、忆阻神经元电路，建立了忆阻自突触耦合神经元模型，为忆阻神经网络系统的实现奠定了基础。设计了基于随机忆阻的神经网络、基于忆阻神经突触的混沌神经网络电路及对双极值图像的联想记忆应用、面向图像复原的忆阻连续Hopfield神经网络电路。研究了忆阻神经网络全电路实现方式，主要包括：两层忆阻神经网络、Transformer网络、忆阻自编码器和忆阻脉冲神经网络。特别在长短时记忆网络的电路实现方面，设计了基于忆阻的短时记忆向长时记忆自动转换电路，探讨了基于生物长短时记忆原理的眼睛疲劳状态检测电路。研究了基于忆阻器控制电路的多涡卷混沌系统、无平衡点三维混沌系统、基于忆阻器的多样化隐藏吸引子混沌系统、基于忆阻器的多双涡卷蔡氏混沌系统，基于移位控制器的Jerk系统的电路实现。对基于忆阻多双涡卷Hopfield神经网络模型进行了动力学分析及电路设计，同时也对基于三阶、四阶忆阻电路的切换网络进行了稳定性分析。设计了基于忆阻的联想记忆泛化和分化电路、基于忆阻的双向联想记忆电路，构建了基于忆阻的操作性条件作用和经典联想记忆级联电路，给出了情感多联想学习忆阻电路，实现了基于忆阻混沌神经网络的动态联想记忆。.依托本项目，共培养6名博士生和45名硕士生，出版修订教材1本，授权国家发明专利4项，举办2次国际学术会议、3次国内学术会议，发表论文47篇，其中ESI高被引论文3篇，ESI热点论文1篇，IEEE 汇刊论文16篇，SCI（或源刊）收录30篇，被他人引用450余次，论文主要发表IEEE T. on Circuits and Systems I: Regular Papers, IEEE T. on Industrial Electronics, IEEE T. on Cybernetics, IEEE T. on Fuzzy Systems, IEEE T. on Biomedical Circuits and Systems等刊物上。