基于神经网络的时序信息识别研究
发布时间：2018-09-01 09:32
【摘要】：神经网络一直以来都是学术界研究的热点,而伴随着图形硬件的更新换代,目前基于深度学习的神经网络再次在各个领域取得丰硕成果。然而这些人工神经网络处理信息时并没有完整的考虑到生物神经元的运行机制,而作为神经科学领域最新研究成果的Spiking神经网络具有高度的生物仿真性,其能够很好的处理时空域维度上的特性信息,并将外界刺激以时间为特征进行编码,最后将编码后的脉冲信息传入神经系统进行处理,在生物身份认证、语音识别等领域取得了很多的实际应用成果。大脑生物神经元能够识别具有脉冲特性刺激信息的时序,然而这种识别机制的原理并没有得到很好的揭示。研究基于Spiking神经网络的时序信息识别能更深入了解大脑对信息处理的原理,从而可将其识别原理应用到识别处理外界复杂时空特性的信息,因此本文的研究非常具有科研前景。总的来说本文的研究内容如下:1.介绍了基于Spiking神经网络时序信息识别所涉及的相关基础知识,包括神经元的相关生物特性和脉冲神经网络模型。同时从信息编码、学习神经元的训练和解码时序神经元模型的构建三大模块进行相关理论的阐述。2.提出了一种新的监督学习算法DL-PSD。针对脉冲神经网络中的时序信息识别,在经典算法PSD基础上,结合神经元的时间延迟特性提出了DLPSD算法,提高了时序信息识别中项识别的效率。3.提出了一种改进的脉冲神经网络序列解码机制,构建相应的神经元模型。传统基于卷积的方式对时序信息的识别并没有充分利用神经元的生物特性,本文根据FSA识别手写字的基本原理,结合了生物神经元树突存在双稳态平台电压的生物特性,构建一种解码特定时序的解码单元模型4.最后将信息编码、项识别以及时序信息解码三大模块构成一个整体进行时序信息的识别。相位编码转换图像信息、DL-PSD训练学习神经元完成项识别、新的解码结构模型识别特定的数字图像序列。实验成功识别了特定的光学字符输入序列,同时改变学习输出神经元与解码模型感知单元的连接结构,可以识别出更多的特定数字序列,展示本文识别机制的可扩展性和鲁棒性,这对于构建通用的识别结构来编码和处理人体生物特征信息提供了一条新的途径,在图像处理领域也有其应用价值。
[Abstract]:Neural network has always been a hot topic in academic research. With the upgrading of graphics hardware, the neural network based on in-depth learning has once again achieved fruitful results in various fields. However, these artificial neural networks do not take into account the operation mechanism of biological neurons when processing information, and Spiking neural networks, as the latest research results in the field of neuroscience, have a high degree of biological simulation. It can deal with the characteristic information in the spatial dimension well, and encode the external stimulus with the characteristic of time. Finally, the encoded pulse information afferent neural system can be processed and authenticated in the biological identity. Speech recognition and other fields have achieved a lot of practical results. Brain biological neurons can recognize the timing of impulsive stimuli, but the principle of this recognition mechanism has not been well revealed. The research of time series information recognition based on Spiking neural network can better understand the principle of brain information processing, so it can be applied to the recognition and processing of complex temporal and spatial information. Therefore, the research of this paper is very promising. In general, the contents of this study are as follows: 1. This paper introduces the basic knowledge of temporal information recognition based on Spiking neural network, including the related biological characteristics of neurons and the model of impulsive neural network. At the same time, from the information coding, learning neuron training and decoding time series neuron model construction of three modules to explain the relevant theory. 2. A new supervised learning algorithm, DL-PSD., is proposed. In this paper, based on the classical algorithm PSD and the time delay characteristic of neurons, a DLPSD algorithm is proposed for the recognition of temporal information in impulsive neural networks, which improves the efficiency of item recognition in time series information recognition. An improved sequence decoding mechanism based on impulse neural network is proposed and the corresponding neuron model is constructed. The traditional method based on convolution does not make full use of the biological characteristics of neurons. According to the basic principle of FSA recognition and writing, this paper combines the biological characteristics of biological neuron dendrites with bistable plateau voltage. A decoding unit model for decoding specific timing is constructed. Finally, three modules, namely information coding, item recognition and timing information decoding, are integrated to recognize the timing information. DL-PSD trains learning neurons to complete item recognition, and a new decoding structure model is used to recognize specific digital image sequences. The experiment successfully identified a specific optical character input sequence, and changed the connection structure between the learning output neuron and the decoding model perception unit, so that more specific digital sequences could be identified. It shows the extensibility and robustness of the recognition mechanism in this paper, which provides a new way to code and process the biometric information of human body by constructing a universal recognition structure, and also has its application value in the field of image processing.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP183
本文编号：2216779
[Abstract]:Neural network has always been a hot topic in academic research. With the upgrading of graphics hardware, the neural network based on in-depth learning has once again achieved fruitful results in various fields. However, these artificial neural networks do not take into account the operation mechanism of biological neurons when processing information, and Spiking neural networks, as the latest research results in the field of neuroscience, have a high degree of biological simulation. It can deal with the characteristic information in the spatial dimension well, and encode the external stimulus with the characteristic of time. Finally, the encoded pulse information afferent neural system can be processed and authenticated in the biological identity. Speech recognition and other fields have achieved a lot of practical results. Brain biological neurons can recognize the timing of impulsive stimuli, but the principle of this recognition mechanism has not been well revealed. The research of time series information recognition based on Spiking neural network can better understand the principle of brain information processing, so it can be applied to the recognition and processing of complex temporal and spatial information. Therefore, the research of this paper is very promising. In general, the contents of this study are as follows: 1. This paper introduces the basic knowledge of temporal information recognition based on Spiking neural network, including the related biological characteristics of neurons and the model of impulsive neural network. At the same time, from the information coding, learning neuron training and decoding time series neuron model construction of three modules to explain the relevant theory. 2. A new supervised learning algorithm, DL-PSD., is proposed. In this paper, based on the classical algorithm PSD and the time delay characteristic of neurons, a DLPSD algorithm is proposed for the recognition of temporal information in impulsive neural networks, which improves the efficiency of item recognition in time series information recognition. An improved sequence decoding mechanism based on impulse neural network is proposed and the corresponding neuron model is constructed. The traditional method based on convolution does not make full use of the biological characteristics of neurons. According to the basic principle of FSA recognition and writing, this paper combines the biological characteristics of biological neuron dendrites with bistable plateau voltage. A decoding unit model for decoding specific timing is constructed. Finally, three modules, namely information coding, item recognition and timing information decoding, are integrated to recognize the timing information. DL-PSD trains learning neurons to complete item recognition, and a new decoding structure model is used to recognize specific digital image sequences. The experiment successfully identified a specific optical character input sequence, and changed the connection structure between the learning output neuron and the decoding model perception unit, so that more specific digital sequences could be identified. It shows the extensibility and robustness of the recognition mechanism in this paper, which provides a new way to code and process the biometric information of human body by constructing a universal recognition structure, and also has its application value in the field of image processing.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP183
【参考文献】
相关期刊论文 前3条
1 蔺想红;王向文;张宁;马慧芳;;脉冲神经网络的监督学习算法研究综述[J];电子学报;2015年03期
2 李婷;;国内外神经网络的发展及概述[J];知识经济;2013年18期
3 曹平;陈盼;章文彬;张潮;;基于脉冲神经网络的语音识别方法的初步探究[J];计算机工程与科学;2008年04期
相关硕士学位论文 前1条
1 沈虹;Spiking神经网络及其应用研究[D];南京工业大学;2006年
，本文编号：2216779
本文链接：https://www.wllwen.com/kejilunwen/zidonghuakongzhilunwen/2216779.html
