CASIA OpenIR  > 毕业生  > 博士学位论文
学习和记忆环路的多尺度脑神经计算模拟
张铁林
学位类型工学博士
导师徐波 ; 曾毅
2016-05
学位授予单位中国科学院大学
学位授予地点北京
学位专业模式识别与智能系统
关键词学习和记忆 生物网络
其他摘要
    学习和记住身边出现的各种信息是大多数哺乳动物所共有的关键认知功能之一。我们有时可以轻松地记录下庞大的视频、图像等信息,但有时又需要付出极大的努力才能搞清哪怕一个英语单词的含义。大脑内部的记忆认知机理是怎样的?从单个的神经元,到成百上千的神经元组成的神经微环路,进而到主管不同记忆类型的神经脑区,不同的结构和组织又表征了怎样的记忆功能?本文将从以下几方面,对记忆的编码、存储、提取等功能进行多尺度的神经环路分析和建模,并将受生物启发的实验结果应用于尖峰神经网络和卷积神经网络的智能学习训练中,使得新构建的网络能够在“机理上类脑”和“功能上类人”。本文的创新点可以总结为以下几个部分:
    (1)创新性地实现了从鼠脑介观尺度到微观尺度数据的网络转化和预测。现有的鼠脑数据中,宏观和介观数据众多,但是唯独缺少对精细建模十分重要的微观尺度数据。基于此,将根据亚区之间的连接状况来预测亚区内部的微观连接特性,如亚区内Motif分布比例、亚区内神经元和突触的数量等,此工作将为建立鼠脑微观结构脑模型提供数据基础。
    (2)创新性地实现了鼠脑微观网络的放电模式(功能)和突触连接(结构)之间的相互预测。现有方法采集得到的生理实验数据,或者偏重于网络功能放电信息,或者偏重于网络连接结构信息,大体量的两者信息的同时获取还暂未实现。本文将尝试采用模型预测的方法解决该类问题:如采用传播激活方法结合结构信息来预测放电Spike信息;采用“时序”和“共现”两种方法结合放电Spike信息来预测连接结构信息。此工作将为构建动态鼠脑模型提供结构和功能基础。
    (3)创新性地采用生物神经元模型、生物网络结构连接规则来构建较大规模的海马区网络模型。现有的与记忆高度相关的海马区生物模型中,大多停留在生物现象的模拟层面,同时具有详实的生物基础和较好的实际场景应用特性的模型还很少。本文通过对多尺度的生物网络结构和功能的模拟,构建具有生物基础的海马区模型。该模型不仅仅具有类生物的学习、记忆功能,更有明显的噪声抑制功能,且能在实际任务评测如MNIST手写体数字识别、赛道数据识别任务中展现出相比传统的深度网络方法更大的抗噪优势。
    (4)创新性地将总结得到的7条生物启发学习规则应用到现有的尖峰神经网络的学习训练等分类任务中,规则包括神经元的生长和消亡、突触的可塑性变化、网络背景噪声分布、时序依赖的突触可塑性、兴奋性抑制性细胞的配比等多种规则。实验结果证明,随着越来越多的生物启发机制的加入,网络将取得越来越高的分类精度。
   (5)创新性地将高层生物全局结构信息的反馈机制加入到卷积神经网络(Convolutional Neural Network,CNN)中。提出HCNN(HOG improved CNN)模型,该模型从结构和原理上更加地类人,且处理能力上也与人高度相似,如模型能在保持对自然图片识别精度的基础上,明显地提高对非自然图像(如具有某种类别图像的纹理特征且杂乱分布)的识别精度。
;
    Learning and remembering the various of information around us is one of the key cognitive functions for most of mammalian animals. Sometimes we can save a large amount of knowledge about videos and images, but sometimes we have to pay much of attentions and efforts for a new single English word. What is the cognitive mechanism of memory in our brain? From a single neuron at microscale level, to the thousands of neurons and neural circuits at mesoscale level, even to the functional brain regions at macroscale level, what is the relationship between structures and functions in these multi-scale data? In this paper, some attempts will be achieved for the analysis of memory coding, storage and retrieval. In addition, some biological inspired rules will be integrated into the training procedure of spiking neural networks and convolutional neural networks in order to realize the performance of human-like intelligence. The main contribution of this paper could be listed as below:
    (1) Provide a mathematical model to convert the neural network information from mesoscale to microscale. The microscale data is few but more important for the detailed network implementation than data from other two scales. So we will try to get the the number of neurons, the number of synapses, the connectivity characteristics within a region (e.g. Motif distribution) by the connectivity data from mesoscale. This work will be the important structural bases of the microscale brain models.
    (2) Make the prediction of neural functions by neural connections in microscale mouse brain, or vise versa. The anatomical and biophysical experiments nowadays are more focus on only neural functions (e.g. spikes) or neural connections (e.g. synaptic connections) individually. In this section, we will try to resolve this kind of problems by proposing two converting models: the spike prediction model based on spreading activation method; the neural connectivity prediction based on “co-occurrence method” and “time-sequence method”. This work will be the important structural and functional bases for dynamical mouse brain models in next steps.
    (3) Integrate the biological neuron models, the rules inspired from biological experiments to build the large scale hippocampus network model. Most of the hippocampus models until now are more focus on the simulation of biological mechanisms, and seldom could meet both the solid biological background and fine functional performance in practice. We build a biological hippocampus network by multi-scale network construction, which not only shows the characteristics of biological learning and memory, but also shows abilities of anti noise. Experimental results show the better performance of proposed model than traditional neural networks (e.g. convolutional neural networks).
    (4) Add seven biological inspired rules into the learning and training procedure of spiking neural networks, such as the formation and elimination of neurons and synapses, the distribution of background noises, the spike timing dependent plasticity rules, the proportion of excitatory and inhibitory neurons. The experimental results show that, with more rules integrated into the network, a better classification performance will be achieved.
    (5) Add the feedback mechanisms from higher layers into traditional convolutional network, and resolve the “fooling images problems” of convolutional neural network by proposed HCNN model (HOG improved CNN). The results show that the new model could correctly distinguish fooling images (in which some local features form the natural images are chaotically distributed) from natural images, and at the same time, not decrease the performance of natural image classification tasks.
学科领域类脑智能
语种中文
文献类型学位论文
条目标识符http://ir.ia.ac.cn/handle/173211/11657
专题毕业生_博士学位论文
作者单位中国科学院自动化研究所
推荐引用方式
GB/T 7714
张铁林. 学习和记忆环路的多尺度脑神经计算模拟[D]. 北京. 中国科学院大学,2016.
条目包含的文件
文件名称/大小 文献类型 版本类型 开放类型 使用许可
张铁林博士论文.pdf(17481KB)学位论文 暂不开放CC BY-NC-SA请求全文
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[张铁林]的文章
百度学术
百度学术中相似的文章
[张铁林]的文章
必应学术
必应学术中相似的文章
[张铁林]的文章
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。