脉动型神经元网络的联想记忆与分割

以广泛讨论的Hodgk in-Hux ley神经元节点组成脉动神经元网络,从神经系统空时模式编码理论研究网络的记忆(或模式)存储与时间分割问题。给定一个输入模式,它是几种模式的叠加,网络能够以一部分神经元同步发放的形式一个接一个地在时间域分割出每一种模式。如果输入的模式是缺损的,系统能够把它们恢复成完好的原型,即神经网络的联想记忆功能。     

基于Hindmarsh-Rose神经元模型的模式分割

联想记忆和模式分割是大脑的重要功能,大脑海马区是这一功能的重要物质基础。Hindmarsh-Rose神经元模型能够较好地模拟大脑海马区神经元的特性。本文以Hindmarsh-Rose神经元为节点,按Hopfield规则构造网络,同时,考虑到大脑中神经元连接的生理学实际情况,通过数值模拟,实现了在弱、中、强三种连接强度情况下的模式分割以及在随机概率连接情况下的模式分割,通过分析比较,提出了进一步改进的方向。     

单调激活函数惯性项神经耦合系统的混沌共存

混沌及其稳态共存是神经网络系统中一个重要研究热点问题．本文基于惯性项神经元模型,利用非线性单调激活函数构造了一个惯性项神经耦合系统,采用理论分析和数值模拟相结合的方法,研究了系统平衡点以及静态分岔的类型,分析了系统两种不同模式的混沌及其稳态共存．具体来说,我们通过选取不同的初始值,利用相应的相位图和时间历程图,展现了系统混沌对初值的敏感依赖性．进一步,采用耦合强度作为动力学的分岔参数,研究了混沌产生的倍周期分岔机制,得到了单调激活函数耦合下的惯性项神经元系统混沌共存现象．     

神经网络的分叉理论设计方法

本文用分叉理论的规范形方程设计和综合期望贮存静、动态记忆模式的神经网络。对于期望贮存静态记忆模式的网络，该规范形方程为叉形分叉的；若期望贮存的记忆模式是周期振荡形式，该规范形方程为高余维数Ｈｏｐｆ分叉的，由满足设计约束的规范形系数得到的突触连接系数可以保证期望贮存的记忆模式都能成功地存贮于所设计的网络，且是网络仅有的吸引子，没有伪吸引子，吸引域的范围足够大。     

基于LSTM-GAN的无人机飞行数据异常检测算法

飞行数据的异常检测是保障无人机飞行安全的关键。为了提升异常检测的准确率,以更好地构建无人机飞行数据正常样本的数据模式,提出了一种基于长短时记忆网络（LSTM）和生成对抗网络（GAN）的异常检测算法。设计了由生成网络、判别网络和重构网络组成的循环学习网络,同时为了避免网络学习产生梯度爆炸的风险,设计了一种由两个判别损失函数与一个重构损失函数相结合的目标损失函数。实验结果表明,LSTM-GAN异常检测算法均优于K-means、单类支持向量机、LSTM和Auto-Encoder算法的异常检测性能。LSTM-GAN比LSTM异常检测的准确率提高2.2%。     

基于神经网络的多种油样分析技术融合诊断

针对4种最常用的油样分析技术(铁谱、光谱、颗粒计数及理化指标分析)的信息融合问题，依据基于规则的专家系统方法，建立了各油样分析技术的子专家诊断系统；依据专家经验建立各子专家系统的诊断结果与故障论域中各故障模式的关系，得到了用于神经网络学习的训练样本．在此基础上，通过对神经网络进行训练，并将待分析油样的各子专家系统诊断结论输入训练成功的网络，即得到融合诊断结果．实例分析表明所建立的分析方法便捷有效．     

平衡特性化联想神经网络的动态特性研究

在反馈联接的联想神经网络中，学习过程是通过神经元联拉突触及其状态的彝实现的，而学习的模式则必须由神经元稳定的状态亚在而网络的全局稳定就涉在触动态系统和神经元活化的联合平衡。本文首先分析学习反馈联想神经网络可以当作时不变系统处理的条件，指出基于平衡点特性化的联想网络的学习速率约束，导出了慢突触变化联想网络平衡点稳定的约束条件，并给出网络在结构扰动情况下网络平衡点运动轨这界的估计。     

一类忆阻神经元的电活动多模振荡及Hamilton 能量反馈控制

根据法拉第电磁感应定律,在离子穿越细胞膜或者在外界电磁辐射下,细胞内外的电生理环境会产生电磁感应效应,继而会影响神经元的电活动行为. 基于此,本文考虑电磁感应影响下的 Hindmarsh-Rose (HR) 神经元模型,研究了其混合模式振荡放电特征,并设计一个 Hamilton 能量反馈控制器,将其控制到不同的周期簇放电状态. 首先,通过理论分析发现磁通 HR 神经元系统的 Hopf 分岔使其平衡点的稳定性发生了改变,并产生极限环,进而研究了 Hopf 分岔点附近膜电压的放电特征. 基于双参数数值仿真发现该系统具有丰富的分岔结构,在不同的参数平面上存在倍周期分岔、伴有混沌的加周期分岔、无混沌的加周期分岔以及共存的混合模式振荡. 最后,为了有效控制膜电压的混合模式振荡,利用亥姆霍兹理论计算出磁通 HR 神经元系统的 Hamilton 能量函数并设计 Hamilton 能量反馈控制器,通过数值仿真分析了膜电压在不同反馈增益下的簇放电状态,发现该控制器能够有效地控制膜电压到不同的周期簇放电模式. 本文的研究结果为探究电磁感应下神经元的分岔结构及其能量控制领域提供了有用的理论支撑.      

一种用于陀螺随机漂移预测的多尺度混合建模方法

针对陀螺随机漂移时间序列由于非平稳和非线性造成单一预测模型难以准确跟踪其变化趋势的问题,提出了一种基于集合经验模态分解(EEMD)和灰色极端学习机(GELM)的多尺度混合建模方法。首先,利用集合经验模态分解将随机漂移时间序列按照频率高低分解为多个本征模式分量和一个余量;然后针对不同类型时频特性分量选择合适激活函数和隐层神经元数目的GELM分别进行预测;最后,以等权相加的方式得到最终预测结果。将该方法用于某型激光陀螺随机漂移预测中,仿真结果表明:混合预测模型能够准确预测陀螺随机漂移,预测精度比残差GM(1,1)和GELM预测模型分别提高了33.43%和23.47%,可为激光陀螺的漂移补偿、故障预报和可靠性诊断提供依据。     

MULTI-MODE OSCILLATIONS AND HAMILTON ENERGY FEEDBACK CONTROL OF A CLASS OF MEMRISTOR NEURON$^{\bf 1)}$

根据法拉第电磁感应定律,在离子穿越细胞膜或者在外界电磁辐射下,细胞内外的电生理环境会产生电磁感应效应,继而会影响神经元的电活动行为. 基于此,本文考虑电磁感应影响下的 Hindmarsh-Rose (HR) 神经元模型,研究了其混合模式振荡放电特征,并设计一个 Hamilton 能量反馈控制器,将其控制到不同的周期簇放电状态. 首先,通过理论分析发现磁通 HR 神经元系统的 Hopf 分岔使其平衡点的稳定性发生了改变,并产生极限环,进而研究了 Hopf 分岔点附近膜电压的放电特征. 基于双参数数值仿真发现该系统具有丰富的分岔结构,在不同的参数平面上存在倍周期分岔、伴有混沌的加周期分岔、无混沌的加周期分岔以及共存的混合模式振荡. 最后,为了有效控制膜电压的混合模式振荡,利用亥姆霍兹理论计算出磁通 HR 神经元系统的 Hamilton 能量函数并设计 Hamilton 能量反馈控制器,通过数值仿真分析了膜电压在不同反馈增益下的簇放电状态,发现该控制器能够有效地控制膜电压到不同的周期簇放电模式. 本文的研究结果为探究电磁感应下神经元的分岔结构及其能量控制领域提供了有用的理论支撑.     

Pattern Recognition and Synchronization in Pulse-Coupled Neural Networks

This paper establishes and studies equations of a network that is composed of neurons with their dendrites and axons. The pulse generation in the axon is described by means of phase oscillators, whereas the dendritic currents are described by linear damping equations with source terms as which the incoming pulses act. The equations take into account delays and noise. In the case of phase-locking the equations have been solved previously by the author. The main objective of the present paper is to study in how far these equations may serve for pattern recognition. It is shown that either a random phase approximation for the stored patterns or a triple interaction term suffice to treat pattern recognition. It is pointed out how then recognized patterns can be either encoded as completed prototype patterns or as phase-locked states. It is suggested that to understand pattern recognition it is not only necessary to consider the whole network instead of “grandmother” cells but equally well the whole intermediate steps.     

Parameter-dependent synchronization of coupled neurons in cold receptor model

In this paper, synchronization in two coupled neurons with spiking, bursting and chaos firings is investigated as the coupling strength gets increased. Synchronization state can be identified by means of the bifurcation diagram, the correlation coefficient and ISI-distance. It is illustrated that the coupled neurons can exhibit different types of synchronization state when the coupling strength increases. The different synchronization processes appear similar, but their detailed processes are different depending on the parameter values. The synchronization of neuronal network with two different network connectivity patterns is also studied. It is shown that chaotic and high period pattern are more difficult to get complete synchronization than the situation in single spike and low period pattern. It is also demonstrated that the synchronization status of multiple neurons is dependent on the network connectivity patterns. These results may be instructive to understand synchronization in neuronal systems.     

Turing patterns of Gierer–Meinhardt model on complex networks

Gierer–Meinhardt (G–M) model is a classical reaction diffusion (RD) model to describe biological and chemical phenomena. Turing patterns of G–M model in continuous space have attracted much attention of researchers. Considering that the RD system defined on discrete network structure is more practical in many aspects than the corresponding system in continuous space, we study Turing patterns of G–M model on complex networks. By numerical simulations, Turing patterns of the G–M model on regular lattice networks and several complex networks are studied, and the influences of system parameters, network types and average degree on pattern formations are discussed. Furthermore, we present an exponential decay of Turing patterns on complex networks, which not only quantitatively depicts the influence of network topology on pattern formations, but also provides the possibility for predict pattern formations.

     

基于BP神经网络的条纹级数插值方法

数据采集是光测力学数据处理自动化的关键和难点，特别是对条纹稀疏和低级数区．针对上述困难，本文利用ＢＰ神经网络能对函数进行逼近的特征，对条纹插值进行了探讨；经研究改进后的算法，其收敛速度明显提高．作为例子，给出了一幅光弹性等差线条纹图的插值实验，获得了令人满意的结果．     

Design of Coupled Nonlinear Systems for Storage of Prescribed Binary Patterns

In this paper, we investigate synchronization and cluster formation phenomena in two-dimensional arrays of locally interconnected chaotic circuits. We report the existence of an abundance of attractors, for which each cell stores a binary information. We describe a simple method for storing binary patterns in the network. We also address the question which patterns can be successfully stored in the network and discuss problems of pattern stability and influence of parameters mismatch. This research has been supported in part by the European Community research program “COSYC of SENS”, no. HPRN-CT-2000-00158 and by AGH-UST grant 11.11.120.182.     

Efficient implementation of the proper outlet flow conditions in blood flow simulations through asymmetric arterial bifurcations

We present an efficient implementation of the proper (in vivo) outlet boundary conditions in detailed, three‐dimensional (3D) and time‐periodic simulations of blood flow through arteries. This is achieved through the intermediate use of an approximate ‘simulant’ model of the outlet pressure/flow relationship corresponding to the full 3D and time‐dependent numerical simulation. This model allows us to efficiently couple the 3D outlet pressure/flow conditions to the equivalent relations due to the downstream arterial network, as obtained from a one‐dimensional approximate model in the form of Fourier frequency impedance coefficients. An adjustable time‐periodic function correction term in the simulant model requires input from the full 3D model that has to run iteratively until convergence. The advantage of the proposed numerical scheme is that it decouples the upstream detailed simulation from the downstream approximate network model offering exceptional versatility. This approach is demonstrated here in a series of detailed 3D simulations of blood flow, performed using the commercial software FLUENT?, through an asymmetric arterial bifurcation. Two cases are considered: first a healthy system patterned after the left main coronary arterial bifurcation, and second a diseased case where an occlusion has developed in one of the daughter vessels, resulting in strengthening the asymmetry of the bifurcation. Rapid convergence of the iterative process was achieved in both cases. Subtle changes occur in the shear patterns of the daughter vessels, whereas the flow distribution is quite different. In the presence of a stenosis additional regions of low shear develop due to inertial effects. Copyright © 2010 John Wiley & Sons, Ltd.     

Turing instability and pattern formation of neural networks with reaction–diffusion terms

In this paper, a model for a network of neurons with reaction–diffusion is investigated. By analyzing the linear stability of the system, Hopf bifurcation and Turing unstable conditions are obtained. Based on this, standard multiple-scale analysis is used for deriving the amplitude equations of the model for the excited modes in the Turing bifurcation. Moreover, the stability of different patterns is also determined. The obtained results enrich the dynamics of neurons’ network system.     

Arm motion control model based on central pattern generator

According to the theory of Matsuoka neural oscillators and with the consideration of the fact that the human upper arm mainly consists of six muscles, a new kind of central pattern generator(CPG) neural network consisting of six neurons is proposed to regulate the contraction of the upper arm muscles. To verify effectiveness of the proposed CPG network, an arm motion control model based on the CPG is established.By adjusting the CPG parameters, we obtain the neural responses of the network, the angles of joint and hand of the model with MATLAB. The simulation results agree with the results of crank rotation experiments designed by Ohta et al., showing that the arm motion control model based on a CPG network is reasonable and effective.     

神经系统信息处理和异常功能的复杂动力学

神经系统通过电活动实现信息处理及生物功能,电活动的节律和时空行为是功能的动力学表征.神经电生理实验结合理论模型,借助于分岔揭示了外界激励、参数和噪声调控下的周期、混沌和随机等多样性的节律模式及其节律的复杂转迁规律,揭示了感觉神经对信息(如血压压力信号和痛觉信息)的节律编码机制,揭示了突触噪声扩大脑神经元的信息传递能力并对能力强弱进行了分类,结果可用于提高信息检测能力和指导镇痛;借助于单神经元节律的动力学——如分岔和簇放电节律的快慢动力学——解释了网络功能异常的时空行为,如药物调控脑皮层的螺旋波/癫痫和慢抑制耦合调控的运动网络的同步转迁/运动模式异常,结果给出了调控系统功能的途径;通过大数据分析获得自闭症患者的脑功能网络的时空行为特征——症状相关脑区的同步活动降低,给出了用于诊断的潜在指标.通过新实验发现、新建理论模型、新分析方法和新观点阐释,揭示了神经系统的复杂动力学,认识和解释了神经系统的信息处理机制和异常生物功能/疾病,具有重要科学意义和潜在应用价值.