新型脉冲神经网络及其相变材料实现的仿真
文献类型：学位论文
| 作者 | 王昊 |
| 学位类别 | 硕士 |
| 答辩日期 | 2013 |
| 授予单位 | 中国科学院上海光学精密机械研究所 |
| 导师 | 阮昊 |
| 关键词 | 脉冲神经网络，相变材料，动态脉冲响应神经元模型DSRM，无监督STDP学习方式，胜者为王 |
| 其他题名 | New Type Spiking Neural Network and Completion Simulated by Phase-Change Material |
| 中文摘要 | 人工神经网络的研究起于1943年，大规模的神经网络由数量众多的神经元组成，这些神经元彼此之间相互连接，通过一定的算法可以使其具备认知能力。认知能力是神经网络要解决的中心问题，而识别则是神经网络应具备的基础能力。 传统的人工神经网络，诸如BP（Back-Propagation）、RBF（Radial Basis Function）、SVM（Support Vector Machine）等神经网络，已经是很成熟的神经网络，它们在识别和认知方面有着大量成熟的算法，而且现在应用得非常广泛，它们可以很好的解决一些模式识别的问题。而被誉为第三代神经网络的脉冲神经网络Spiking Neural Network（SNN）是建立在真实的生物神经元基础之上的。 本文提出了一种动态脉冲响应神经元模型DSRM（Dynamic Spike Response Model），该神经元模型是以SRM（Spike Response Model）模型为基础而改进的，本文基于DSRM神经元模型构建了具有无监督STDP（Spike-Timing Dependent Plastistiy）学习方式的脉冲神经网络。通过对脉冲神经网络的训练以及输出层神经元之间的竞争使网络具有识别能力。除此之外，本文还介绍了相变材料的电学特性，并利用其电学特性对STDP学习方式进行了仿真。 本文首先阐述了组成脉冲神经网络的神经元的模型，而后详细介绍了构建脉冲神经网络的算法，并使神经网络对英文字符和二维码分别进行识别。通过无监督STDP学习方式对脉冲神经网络进行训练来达到输入神经元与输出神经元之间权重的收敛，输出神经元之间的竞争采用胜者为王的竞争策略使其与待识别模式建立对应关系，在识别过程中权重并不做更新，这使得识别的过程基本可以达到实时。最终把相变材料电学特性仿真的STDP学习过程添加到DSRM神经元模型中。 |
| 英文摘要 | The research of Artificial Neural Network begins from 1943, a large-scaled neural network compsed of hundreds of thousands of neurons can be provided with the ability of cognition through training. The main problem attempting to be solved by neural network is cognition, and the principal capacity of neural network is cognition. Traditional neural networks, such as BP (Back-Propagation), RBF (Radial Basis Function), SVM (Support Vector Machine), in which a large number of mature algorithm can be utilized, have been greatly acceptable in the field of pattern recognition. Spiking Neural Network (SNN), the third generation of neural network, is established in accordance with biological neurons. The present paper uses DSRM (Dynamic Spike Response Model) model which is improved by SRM (Spike Response Model) equipped with STDP (Spike-timing Dependent Plastisity) learning rule to build a Spiking Neural Network. The SNN can have the ability of pattern cognition though training as well as competition of output neurons. Additionally, the paper introduces the electronic characteristic of Phase-Change Material, and simulates the STDP learning rule using this electronic characteristic. We first introduce the improved SRM model, then illustrate the algorithms to establish the network, and finally we use the SNN to identify letters and two-dimension code. The weight between input and output neurons can be convergent through training, and the competition among output neurons utilize winner-takes-all strategy to setup the correspondence with the pattern being identified. In the identification stage, weight is no longer refreshed, by which the SNN can meet the qualification of real-time. Finally, we insert the simulation above into the improved neuron model DSRM. |
| 语种 | 中文 |
| 源URL | [http://ir.siom.ac.cn/handle/181231/16760]  |
| 专题 | 上海光学精密机械研究所_学位论文 |
| 推荐引用方式 GB/T 7714 | 王昊. 新型脉冲神经网络及其相变材料实现的仿真[D]. 中国科学院上海光学精密机械研究所. 2013. |
入库方式： OAI收割
来源：上海光学精密机械研究所
浏览0
下载0
收藏0
其他版本
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。