基于突触可塑性的小世界神经网络的动态特性研究
|
摘要
人工神经网络具有大规模的信息处理和存储能力、良好的自适应性以及很强的学习功能、联想功能和容错功能。动态特性的研究一直是人工神经网络理论研究的重点,主要原因在于人工神经网络的应用都与网络的动态特性有关。目前,神经网络的研究主要是基于层级网络,其拓扑不能模拟真实生物神经网络。小世界网络作为大量真实复杂系统的高度抽象,具有生物神经网络特性。本研究首先构建了小世界网络并基于复杂网路理论选择出适合于小世界网络的最佳参数,进而基于突触可塑性调节机制和小世界网络拓扑构建了小世界神经网络,并从放电特性、突触权重动态特性和复杂网络特性三个方面分析了小世界神经网络的动态特性。实验结果表明:随着时间的增加,小世界神经网络的兴奋性与抑制性神经元放电模式没有改变且神经元的放电时间趋于同步;小世界神经网络中各神经元间的突触权重急剧减小最终趋于稳定;网络的连接减弱且信息传递效率降低,但小世界属性较为稳定。小世界神经网络的动态特性随时间而变化且相互影响:网络的放电同步特性可影响突触权重趋于最小值分布,进而突触权重的动态变化也可影响复杂网络特性。
The artificial neural network has the ability of the information processing and storage, good adaptability, strong learning function, association function and fault tolerance function. The research on the artificial neural network is mostly focused on the dynamic properties due to fact that the applications of artificial neural networks are related to its dynamic properties. At present, the researches on the neural network are based on the hierarchical network which can not simulate the real neural network. As a high level of abstraction of real complex systems, the small world network has the properties of biological neural networks. In the study, the small world network was constructed and the optimal parameter of the small word network was chosen based on the complex network theory firstly. And then based on the regulation mechanism of the synaptic plasticity and the topology of the small world network, the small world neural network was constructed and dynamic properties of the neural network were analyzed from the three aspects of the firing properties, dynamic properties of synaptic weights and complex network properties. The experimental results showed that with the increase of the time, the firing patterns of excitatory and inhibitory neurons in the small world neural network didn't change and the firing time of the neurons tended to synchronize; the synaptic weights between the neurons decreased sharply and eventually tended to be steady; the connections in the neural network were weakened and the efficiency of the information transmission was reduced, but the small world attribute was stable. The dynamic properties of the small world neural network vary with time, and the dynamic properties can also interact with each other: the firing synchronization of the neural network can affect the distribution of synaptic weights to the minimum, and then the dynamic changes of the synaptic weights can affect the complex network properties of the small world neural network.
The artificial neural network has the ability of the information processing and storage, good adaptability, strong learning function, association function and fault tolerance function. The research on the artificial neural network is mostly focused on the dynamic properties due to fact that the applications of artificial neural networks are related to its dynamic properties. At present, the researches on the neural network are based on the hierarchical network which can not simulate the real neural network. As a high level of abstraction of real complex systems, the small world network has the properties of biological neural networks. In the study, the small world network was constructed and the optimal parameter of the small word network was chosen based on the complex network theory firstly. And then based on the regulation mechanism of the synaptic plasticity and the topology of the small world network, the small world neural network was constructed and dynamic properties of the neural network were analyzed from the three aspects of the firing properties, dynamic properties of synaptic weights and complex network properties. The experimental results showed that with the increase of the time, the firing patterns of excitatory and inhibitory neurons in the small world neural network didn't change and the firing time of the neurons tended to synchronize; the synaptic weights between the neurons decreased sharply and eventually tended to be steady; the connections in the neural network were weakened and the efficiency of the information transmission was reduced, but the small world attribute was stable. The dynamic properties of the small world neural network vary with time, and the dynamic properties can also interact with each other: the firing synchronization of the neural network can affect the distribution of synaptic weights to the minimum, and then the dynamic changes of the synaptic weights can affect the complex network properties of the small world neural network.
引文
1Qiu Tie,Luo Diansong,Xia Feng,et al. A greedy model with smallworld for improving the robustness of heterogeneous Internet ofThings. Computer Networks,2016,101(C): 127-143.
2Zhou Guangye,Li Chengren,Li Tingting,et al. Outersynchronization investigation between WS and NW small-worldnetworks with different node numbers. Physica A: StatisticalMechanics and its Applications,2016,457(8): 506-513.
3廖志贤,罗晓曙. 基于小世界网络模型的光伏微网系统同步方法研究. 物理学报,2014,63(23): 98-104.
4于凯. 电磁场对神经元网络同步影响的研究. 天津: 天津大学,2013.
5Yu Haitao,Guo Xinmeng,Wang Jiang,et al. Spike coherence andsynchronization on Newman-Watts small-world neuronalnetworks modulated by spike-timing-dependent plasticity. PhysicaA: Statistical Mechanics and its Applications,2015,419: 307-317.
6Dong Ziqian,Wang Zheng,Xie Wen,et al. An experimental studyof small world network model for wireless networks. Journal ofCyber Security,2015,4: 259-278.
7曲海波,吕粟,张文静,等. 幼儿小世界神经网络节点属性与影响因素的相关性分析. 生物医学工程学杂志,2016,33(5): 931-938,944.
8Guan Jinting,Tang Meishuang,Huang Guangzao,et al. A newsmall-world network model for instant messaging chatnetwork//11th System of Systems Engineering Conference (So SE-2015). San Antonio: IEEE,2016: 1-5.
9Sun R. Complex network evolution model based on nodeattraction. Applied Mechanics & Materials,2014,596(7): 843-846.
10Li Xuefei,Chang Lijun,Zheng Kai,et al. Ranking weightedclustering coefficient in large dynamic graphs. World Wide Web:Internet and Web Information Systems,2017,20(5): 855-883.
11Britz J,Van De Ville D,Michel C M. BOLD correlates of EEGtopography reveal rapid resting-state network dynamics.Neuroimage,2010,52(4): 1162-1170.
12尹宁,徐桂芝,周茜. 磁刺激穴位复杂脑功能网络构建与分析. 物理学报,2013,62(11): 569-576.
13郭磊,王瑶,于洪丽,等. 基于近似熵的磁刺激穴位脑功能网络构建与分析. 电工技术学报,2015,30(10): 31-38.
14Izhikevich E M. Simple model of spiking neurons. IEEETransactions on Neural Networks,2003,14(6): 1569-1572.
15Gkoupidenis P,Schaefer N,Strakosas X A,et al. Synaptic plasticityfunctions in an organic electrochemical transistor. Appl Phys Lett,2015,107(26): 155-159.
16Kleberg F I,Fukai T,Gilson M. Excitatory and inhibitory STDPjointly tune feedforward neural circuits to selectively propagatecorrelated spiking activity. Frontiers in ComputationalNeuroscience,2014,8(4): 53.
17王美丽,王俊松. 基于抑制性突触可塑性的反馈神经回路兴奋性与抑制性动态平衡. 物理学报,2015,64(10): 416-423.
18Mu Junfen,Sun Hexu,Wang Jinhuan,et al. A weighted networkmodel with simultaneous change of node and edge//ControlConference. Australian: IEEE,2014: 2805-2809.
19Jiang J R,Huang H W,Liao J H,et al. Extending Dijkstra’s shortestpath algorithm for software defined networking//NetworkOperations and Management Symposium. Krakow: IEEE,2014: 1-4.
2Zhou Guangye,Li Chengren,Li Tingting,et al. Outersynchronization investigation between WS and NW small-worldnetworks with different node numbers. Physica A: StatisticalMechanics and its Applications,2016,457(8): 506-513.
3廖志贤,罗晓曙. 基于小世界网络模型的光伏微网系统同步方法研究. 物理学报,2014,63(23): 98-104.
4于凯. 电磁场对神经元网络同步影响的研究. 天津: 天津大学,2013.
5Yu Haitao,Guo Xinmeng,Wang Jiang,et al. Spike coherence andsynchronization on Newman-Watts small-world neuronalnetworks modulated by spike-timing-dependent plasticity. PhysicaA: Statistical Mechanics and its Applications,2015,419: 307-317.
6Dong Ziqian,Wang Zheng,Xie Wen,et al. An experimental studyof small world network model for wireless networks. Journal ofCyber Security,2015,4: 259-278.
7曲海波,吕粟,张文静,等. 幼儿小世界神经网络节点属性与影响因素的相关性分析. 生物医学工程学杂志,2016,33(5): 931-938,944.
8Guan Jinting,Tang Meishuang,Huang Guangzao,et al. A newsmall-world network model for instant messaging chatnetwork//11th System of Systems Engineering Conference (So SE-2015). San Antonio: IEEE,2016: 1-5.
9Sun R. Complex network evolution model based on nodeattraction. Applied Mechanics & Materials,2014,596(7): 843-846.
10Li Xuefei,Chang Lijun,Zheng Kai,et al. Ranking weightedclustering coefficient in large dynamic graphs. World Wide Web:Internet and Web Information Systems,2017,20(5): 855-883.
11Britz J,Van De Ville D,Michel C M. BOLD correlates of EEGtopography reveal rapid resting-state network dynamics.Neuroimage,2010,52(4): 1162-1170.
12尹宁,徐桂芝,周茜. 磁刺激穴位复杂脑功能网络构建与分析. 物理学报,2013,62(11): 569-576.
13郭磊,王瑶,于洪丽,等. 基于近似熵的磁刺激穴位脑功能网络构建与分析. 电工技术学报,2015,30(10): 31-38.
14Izhikevich E M. Simple model of spiking neurons. IEEETransactions on Neural Networks,2003,14(6): 1569-1572.
15Gkoupidenis P,Schaefer N,Strakosas X A,et al. Synaptic plasticityfunctions in an organic electrochemical transistor. Appl Phys Lett,2015,107(26): 155-159.
16Kleberg F I,Fukai T,Gilson M. Excitatory and inhibitory STDPjointly tune feedforward neural circuits to selectively propagatecorrelated spiking activity. Frontiers in ComputationalNeuroscience,2014,8(4): 53.
17王美丽,王俊松. 基于抑制性突触可塑性的反馈神经回路兴奋性与抑制性动态平衡. 物理学报,2015,64(10): 416-423.
18Mu Junfen,Sun Hexu,Wang Jinhuan,et al. A weighted networkmodel with simultaneous change of node and edge//ControlConference. Australian: IEEE,2014: 2805-2809.
19Jiang J R,Huang H W,Liao J H,et al. Extending Dijkstra’s shortestpath algorithm for software defined networking//NetworkOperations and Management Symposium. Krakow: IEEE,2014: 1-4.