【作者】 杨静；

【导师】 高隽；

【作者基本信息】 合肥工业大学 ， 信号与信息处理， 2009， 博士

【摘要】 联想记忆网络模拟人脑信息存储及回忆机制，具有对含噪及不完全信息的鲁棒处理能力，因而在人工智能、模式识别等领域获得了广泛的研究与应用。复杂网络关注系统结构与功能之间的关系，是近年来研究复杂系统的新视角、新方法。模拟大脑神经学习机理的联想记忆模型其本质表现为一种复杂的非线性动力学系统，同时，生物脑神经系统中普适存在着复杂网络中典型的小世界效应和无标度特性。因此，从复杂网络角度出发，研究稀疏互联联想记忆模型实现就成为了一种新颖的思路。本文借鉴复杂网络研究结构与功能关系新思想，从网络体系结构角度出发，深入而系统地从理论分析和应用实例两方面进行交替互补研究，着重探讨了神经元间稀疏互联方式对于网络联想记忆性能的影响，并构建了相应的复杂网络体系结构下的稀疏互联联想记忆模型。论文的主要工作及创新点包括：(1)综述了联想记忆神经网络相关研究工作，指出其在生物学建模的合理性及硬件实现时存在的问题，分析了采用复杂网络思想研究稀疏互联联想记忆的可行性，提出了从网络体系结构角度出发，从神经元稀疏互联方式入手，借鉴复杂网络研究思想和生物神经系统中普遍存在的复杂网络性质，展开对稀疏互联联想记忆模型理论及应用两方面进行研究的新的思路和方法。(2)借鉴复杂网络研究背景及生物神经系统本身所具有的广泛稀疏连接的内在特点，研究了一类广义稀疏互联联想记忆网络的实现。它可以将现有的各种基于复杂网络结构体系的稀疏互联联想记忆模型整合入一个统一的框架中，以概率统计分析为手段，研究了具有任意连通度的稀疏互联联想记忆网络的动力学演化行为。(3)鉴于大脑皮层神经元突触连接中有限的代谢能量资源限制，使用信噪比分析方法，研究了传统全互联Hopfield网络基础上，有限连接代价限制条件下，网络最优稀疏互联结构的确定原则，以期在降低网络连接成本的同时，最大限度的维持网络性能。(4)研究了复杂网络小世界体系下的联想记忆实现问题，针对原始小世界网络捷径生成具有随机性，缺乏面向任务的确定性操作的缺陷，借鉴复杂动态网络中和谐统一的混合择优模型构建思想，考虑有限连接代价限制条件下网络结构最优稀疏原则引导的捷径生成方式，提出了一种新的小世界体系结构自适应联想记忆模型。新模型可根据学习任务的实际需求，有目的的选择捷径生成，构建任务自适应的网络结构，有效的实现了联想记忆。(5)模拟人脑功能区核磁共振成像所揭示的无标度特性，考虑神经元突触动态生成时融入复杂网络无标度模型形成中的“马太效应”，提出了一种结构动态择优的无标度联想记忆模型。该模型根据有限连接代价限制条件下网络互联结构最优稀疏原则，定义了节点间亲和度的概念，综合考虑了基于节点度值和节点亲和度共同驱动下的择优连接机制，从而同时具有较高的联想记忆性能及神经生物学背景。更多还原

【Abstract】 Associative Memory modeling human intelligence as information storage and recovery is a hot research issue in neural computing, and receives widely application in the field of artificial intelligence and pattern recognition. Complex network focuses on the relationship between topology and function of the network, and it’s a brand-new method of complex system investigation. Associative Memory model is a dynamic nonlinear complex system in essence, and small-word and scale-free characteristics are a universal phenomenon in biological neural system. Therefore, from these points of view, it’s a novel and feasible manner to study Associative Memory through complex network ideology.In this dissertation, we utilize complex network ideology to research the influence of the neurons’ sparse interconnection style on the network performance, i.e. topology versus function. Starting from network topology, we do detailed theoretical analysis and application instance research on Associative Memory realization.The main research contents and innovative contributions of this dissertation are as follows:(1) Status of the Associative Memory modeling technology is summarized. The problems existed in current models as lacking of biological modeling background and its’ unrealistic VLSI implementation are pointed out. The feasibility of researching sparsely connected Associative Memory through complex network is analyzed, and the new ideology of complementary investigation both theoretical and numerical is realized.(2) To go one step closer to more biological realistic model which displays widely sparely connected architecture and meanwhile possesses complex network property, we study a general sparely connected Associative Memory model using probabilistic approach, and explicit analytical solutions for the transient dynamics of the model with arbitrary connectionism are derived.(3) In view of the limited energy consumptions as in the human brain, we derive optimal synaptic dilution strategy under the constraint of limited energy consumptions on fully connected Hopfield network through signal-to-noise analysis. Such synaptic dilution strategy can maintain the network performance utmost while contributing to the energy saving.(4) A novel Associative Memory model based on small-world adaptive structure is proposed in this paper. Aimed at overcoming the disadvantage of random shortcuts formation of the existing methods, this new model takes the ideology of Harmonious Unifying Hybrid Preferential Model and the optimal synaptic dilution strategy under the constraint of limited energy consumptions both into account. This new model breaks the traditional mean of random rewiring but instead constructs a task-based network structure which is much closer to human brain as possessing small-world architecture and can also achieve better performance than the existing counterparts of the same class. The rationality and validity of the proposed model is validated from great number experiments. (5) In order to imitating the scale-free characteristic discovered in the human brain through fMRI. We propose a new Associative Memory model base on dynamic preferential attachment scale-free structure. The conception of affinity between neurons is defined via the optimal synaptic dilution strategy under the constraint of limited energy consumptions. Then the preferential attachment is done by integrate driven of the neuron degree and affinity. The new model not only possesses scale-free architecture but also achieves better performance than the existing counterparts.更多还原