【作者】 刘辉；

【导师】 宋永端； 韩花丽；

【作者基本信息】 重庆大学 ， 工程硕士（控制工程领域）（专业学位）， 2016， 硕士

【摘要】 手写数字识别技术是利用计算机自动辨认手写阿拉伯数字的方法,通常是文献检索、办公自动化、邮件分拣、银行票据处理等系统的核心和关键。由于识别对象特有的复杂性,现有方法停留在一般的模式识别阶段,并没有很好利用知识进行启发诱导,也没能模拟人脑思维过程。为此探讨一种利用类脑结构解决手写数字识别的方法具有重大的现实意义。液体状态机(liquid state machine,LSM)作为一种典型的类脑模型,为解决手写数字识别问题打开了新思路,其储备池结构作为关键处理单元,直接影响模型的精度。由此本文围绕手写数字识别问题,从仿生网络的基本特征多簇和自组织特性出发对液体状态机进行了改进,构建了自组织分簇和分簇自组织网络。自组织分簇网络是利用周期性的电流输入,神经元的放电频率可控,在对称放电时间依赖的突触可塑性(spike-timing-dependent plasticity,STDP)学习机制下,相邻神经元接收相同信号的连接会不断加强,进而出现分簇特征;分簇自组织网络是利用基于时间窗的皮质层生成算法构建多簇的拓扑结构,在此基础上利用非对称STDP学习定义神经元之间的连接强度。由于脑网络的结构功能与其动力学密切相关,本文对两种网络都进行了动力学分析,表明构建网络的优越性,并对两种基本的放电模式(峰放电和簇放电)自组织演化形成的分簇结构也进行了研究。并利用分簇自组织网络优化储备池结构,引入信号重构问题确定液体状态机各参数的大小,建立改进的液体状态机模型。最后利用美国邮政数据MNIST数据库作为数据来源,将手写数字图片输入转换为一系列的脉冲输入,利用信号重构的思想,分别训练每个输出神经元的连接权值,最终得到识别结果。为了节约计算成本,降低能耗,对图片进行了归一化处理。本方法作为液体状态机在手写数字识别任务的首次尝试比2013年Mass提出的用SNN来解决手写数字识别问题具有更明显的优势。用LSM实现手写数字识别关键在于脉冲序列的分类,最后从一个脉冲序列四分类任务中可以看出本方法还有很大的优化空间,对手写数字识别问题具有较大的潜力。更多还原

【Abstract】 Handwritten numeral recognition technology, which is the core of the usual systems including literature search, office automation, mail sorting and bank-note processing, is an automatic digits recognizing method based on the computer. This method stays in general pattern recognition phase without using knowledge and simulating the human brain thinking. So the study of handwritten numeral recognition technology based on the brain like structure is of great practical significance.As a typical brain like model, liquid state machine(LSM) opens a new way for handwriting digits recognition and the reservoir is the key processing unites to improve the accuracy of LSM. Therefore, centering with handwritten digits recognition, we studied the self-organization of two bionic spiking neural networks with multi-cluster and self-organization as their features to modify the LSM: self-organized multi-clustered neural network and multi-clustered self-organized neural network.Self-organized multi-clustered neural network is based on the controllable firing frequency under periodic current injection and the synapses among neighboring neurons in a similar location which receive the same input tend to be strengthened by the symmetric spike-timing-dependent plasticity(STDP)learning rule, leading to the clustered structure; multi-clustered self-organized neural network is built on the basis of multi-clustered network produced by the Kaiser’s developmental time window algorithm, and then, each cluster synaptic weights are further refined through the asymmetric STDP learning rule. As the brain network and dynamics is closely related, we show the superiority of the two networks by analyzing their dynamics. And the emergence of multi-clustered structure of self-organized neural network with different neuronal firing patters, i.e., bursting or spiking, has been investigated. Then the revised LSM with multi-clustered self-organized neural network is built by parameters setting by a benchmark task named bionic signal reconstruction.At last, the weight of each neuron is trained individually with series of pulse input that is transformed form images of handwritten digits from MNIST database(Mixed National Institute of Standards and Technology database) by signal reconstruction. The images are normalized in order to reduce computation cost. The method proposed has obvious advantages than the first attempt to solve handwritten digits recognition with SNN by Mass in 2013. The key of realizing handwritten digits recognition is classifying the input pulse appropriately, and from this perspective, the model has considerable potentials and space for improvement.更多还原