【作者】 蔺想红；

【导师】 张田文；

【作者基本信息】 哈尔滨工业大学 ， 计算机应用技术， 2009， 博士

【摘要】 最近几年来,人工神经网络的研究重点逐渐转向更具生物真实性的脉冲神经网络。随着神经科学研究和技术的快速进展,很多研究者认为基于脉冲定时的大规模神经网络是脑进行信息处理的基础。然而从神经科学的研究成果出发,对脉冲神经网络进行有效的建模和计算,将面临许多概念和技术上的挑战。本文旨在解决脉冲神经网络的两个基本问题:(1)如何模拟单神经元的脉冲动态特性和由多神经元构成的网络;(2)如何发育和进化大规模脉冲神经网络。结合Hodgkin-Huxley神经元模型的动力学特性与IF (Integrate-and-Fire, IF)神经元模型的解析特性,研究了一种新的二维分段线性脉冲神经元模型。该模型的优点在于既可通过分岔理论对兴奋性系统进行定性描述,又可通过状态变量的解析式对神经元行为进行定量分析。通过详细的分析,发现该模型具有许多一维IF神经元模型所不具有的新的神经计算特性。在实验中,应用该模型模拟了大部分已知皮层神经元的脉冲和簇放电行为。神经计算依赖于由神经元模型构成的大规模网络的有效模拟。本文研究了一种新的可进行精确模拟的指数突触电导IF神经元模型,通过单脉冲激励的突触后电位和多脉冲激励的自发放电统计分析,发现该模型的脉冲反应动态特性与指数突触电导被动膜方程模型接近,而计算效率接近脉冲耦合漏电IF神经元模型。同时构建了指数突触电导IF神经元模型的事件驱动模拟策略,并分别应用事件驱动和时钟驱动模拟策略模拟了由指数突触电导IF神经元构成的大规模随机网络,结果表明:(1)在事件驱动模拟策略中,模拟时间和总的脉冲事件数线性成比例;(2)在不同的模拟策略中,脉冲事件的时间精度会影响网络的神经动态特性。基于编码网络结构和动态特性的人工基因组模型,研究了具有固定长度和可变长度的两类基因片断复制与歧化模型进化基因调控网络。应用分析和模拟技术,发现这两类网络具有和生物基因调控网络相近的结构特征,表现出无标度和小世界的拓扑结构。同时也发现这些网络具有和生物基因调控网络相似的动态特性,以更大的概率运转在有序状态,以更小的概率运转在混沌状态。结果表明基因调控网络的生成方法,特别是基因片段复制与歧化,对于网络结构和动态特性的突现具有重要的作用。越来越多的研究表明发育偏差对于生物体的形成具有重要的作用,自然选择并不是唯一决定进化变化方向的因素。本文研究了一种人工细胞谱系发育的计算模型,并用该模型生成各类不同表现型复杂性的随机生物体,分析了变异算子对发育偏差的作用。结果表明:首先,发育系统存在内在的发育偏差,并随着模型参数的不同而变化;其次,发育偏差随着表现型复杂性和变异算子的不同而变化,基因片段复制与歧化算子影响发育偏差的模式;最后,累积变异的发育偏差影响生物体进化变化的方向,并使表现型的复杂性逐渐增长。通过自然进化得到的脑包含几十亿的神经元和几万亿的神经连接,并表现出复杂的智能行为。受生物脑进化的启迪,研究者开辟了进化神经网络的研究领域。以人工基因组模型为框架描述基因调控网络,用基因表达的动态特性表示细胞命运特化的发育过程,研究了一种进化大规模脉冲神经网络的发育方法。该方法的优点在于可以快速有效地发育生成脉冲神经元、神经连接和突触可塑性。相应的食物采集进化实验突现了以神经驱动的自主智能体的智能行为,并验证了该方法对大规模脉冲神经网络的进化能力。更多还原

【Abstract】 During last few years we have witnessed a shift of the emphasis in the artificialneural network community toward spiking neural networks. Motivated by neurosciencediscoveries, many studies consider large-scale neural networks with spike-timing as anessential component in information processing by the brain. While the study of spikingneural networks as a neuroscience research methodology still faces difficult conceptualand technical challenges, it is a promising and timely endeavor. This thesis motivates toaddress two fundamental issues which are: (1) how spiking dynamics of each neuron andnetworks of spiking neurons are simulated; (2) how large-scale spiking neural networksare developed and evolved.Combining the dynamical property of Hodgkin-Huxley neuron model and the ana-lytical property of integrate-and-fire neuron model, we propose a novel two-dimensionalpiecewise linear spiking neuron model. We show that this framework allows a qualita-tive description of excitable systems through bifurcation theory but also a quantitativeanalysis of neuronal behavior through an explicit analytical representation of the statevariables. A detailed analytical study of the model is presented. The model gives riseto new neuro-computational properties not present in one-dimensional integrate-and-fireneuron models. In experiments, using this model we simulated the spiking and burstingbehavior of known types of cortical neurons.Neural computation relies heavily on the simulation of large-scale networks of neu-ron models. We propose a novel integrate-and-fire neuron model with exponential synap-tic conductances that can be simulated exactly. The postsynaptic potentials and sponta-neous discharge statistics of the new model are compared with those of commonly usedmodels, such as the leaky integrate-and-fire model with instantaneous synaptic interac-tions or the passive membrane equation model with exponential synaptic conductancesin which conductances are explicitly integrated. The proposed model is much closer tothe passive membrane equation model with respect to the spiking response dynamics,while still being nearly as computationally efficient as simple leaky integrate-and-firemodel. Then we present an event-driven simulation strategy for the new model. Usingevent-driven and clock-driven simulation strategies we simulate large-scale random net-works, the results show that (1) the simulation time scales linearly with the total numberof spiking events in the event-driven simulation strategies and (2) the temporal precision of spiking events impacts on neuronal dynamics of network in the different simulationstrategies.Based on a model of network encoding and dynamics called the artificial genome,we propose the fixed-size and variable-size segmental duplication and divergence mod-els for evolving genetic regulatory networks. Using analytic and simulation techniques,we find that the two classes of networks share structural properties with natural tran-scriptional regulatory networks. Specifically, these networks can display scale-free andsmall-world structures. We also find that these networks have higher probability to op-erate in the ordered regime, and lower probability to operate in the chaotic regime. Thatis, the dynamics of these networks is similar to that of natural networks. The resultsshow that the structure and dynamics inherent in natural networks may be in part due totheir method of generation rather than being exclusively shaped by subsequent evolutionunder natural selection.There is a growing recognition that natural selection is not the sole determinant ofthe direction of evolutionary change. Recent work has attempted to show the importanceof developmental bias in shaping organic form. We propose a computational model of adevelopmental system of artificial cell lineages. We use the model to generate randomorganisms with different levels of phenotypic complexity and analyze mutation opera-tors for their effects on developmental bias. We find that developmental bias exists in themodel system and varies with the model parameter. We also show that developmentalbias varies strongly with phenotypic complexity and mutation operator. Finally, we il-lustrate how developmental bias affects the direction of evolutionary modification fromthe manner in which accumulated mutations cause an increase in phenotypic complexity.These results suggest that mutation operators could have strong in?uences on develop-mental bias.Research in evolutionary neural networks is inspired by the evolution of biologicalbrains. Because natural evolution discovered intelligent brains with billions of neuronsand trillions of connections, perhaps evolutionary neural networks can do the same. Us-ing the artificial genome model as a framework for describing genetic regulatory net-works, the dynamics of gene expression can be treated as a model for cell fate speci-fication. We propose a developmental method for evolving large-scale spiking neuralnetworks. The advantage of this method is that it can facilitate fast and efficient de-velopment of spiking neurons, neural connections, and synaptic plasticities. The cor-responding evolutionary experiment shows that the intelligent behavior emergences forthe neurally-driven autonomous agents in a food gathering task. Additionally, it also shows that due to the efficiency of the proposed method, large-scale spiking neural net-works can be easily managed thereby making it suitable for long durational evolutionaryexperiments.更多还原