脉冲神经网络学习算法的研究及其应用

发布时间：2018-08-01 09:38

【摘要】：为深入研究生物大脑处理信息以及学习的能力,研究者们提出了人工神经网络,用来模仿大脑信息表达以及处理的过程,而其中具有最高仿生性的是脉冲神经网络,它表达信息以及处理信息均是采用对时间编码的方式。比起感知机等传统神经网络,脉冲神经网络与生物大脑神经元在信息处理机制方面更加接近。许多研究均表明,脉冲神经网络无论是在信息表达能力还是计算能力与传统神经网络相比都更胜一筹。因而它引起了国内外学者的广泛关注和高度重视。目前,脉冲神经网络在人工智能等多方面领域已经有一些初步研究成果,但是远没达到商用的程度,相对传统神经网络等它在实际应用中还是较少的。首先是因为时间先后因素,研究相对并不是那么深入,也还没有普及;再者,虽然脉冲神经网络被证实是与生物神经系统最接近的网络,但是其生物大脑神经系统的学习机制尚不清晰,对网络中神经元学习训练过程的研究也不成熟,因此,学习方法的研究目前仍然是一个值得研究的问题。为了充分运用脉冲神经网络的优点,高度仿生性、较强的信息表达能力以及计算能力,本文对脉冲神经网络监督学习算法进行了深入的研究。目前已经存在一些监督学习算法,但是学习效率或者是适用性等方面还是不够好,为了提高脉冲神经网络的学习效率、精确度以及能够适应更加复杂的问题,本文结合ReSuMe算法、SpikeProp算法等经典算法或者规则,对多层网络监督学习算法进行了优化以及创新,并且对算法进行了仿真和实验。本论文工作内容如下:1)首先分析了目前存在的一些监督学习算法性能、精确度等方面的优缺点,比如SpikeProp算法、ReSuMe算法等。2)提出一种多脉冲多层的神经网络监督学习算法。该算法是结合ReSuMe算法,对目前存在的多层算法进行优化和创新,最后对算法进行了仿真。3)在此基础上,提出了基于延迟的神经网络监督学习算法。该算法使得神经网络学习过程不再单单只是局限于突触权重的调整,对原有的算法进行了扩展与创新,最后对算法进行了仿真。4)最后将此算法成功应用到XOR逻辑异或、Iris数据集分类等问题中,表现出了很好的效果。
[Abstract]:In order to study the ability of processing information and learning in the biological brain, researchers have proposed artificial neural networks, which are used to mimic the process of brain information expression and processing, and the most bionic is pulse neural network. It expresses information and processing information in a time coded way. The neural network, the pulse neural network and the biological brain neuron are closer to the information processing mechanism. Many studies have shown that the pulse neural network is better than the traditional neural network in both the information expression ability and the computing power. Therefore, it has aroused wide attention and attention of scholars at home and abroad. The pulse neural network has some preliminary research results in the field of artificial intelligence, but it is far from commercial. It is still less than the traditional neural network in practical application. First, because of the time successively, the research is relatively not so deep and has not been popularized; moreover, although the pulse neural network is not so popular. The network is proved to be the closest network to the biological nervous system, but the learning mechanism of the neural system of the biological brain is still not clear, and the study of the learning and training process of the neuron in the network is not mature. Therefore, the study of the learning method is still a problem worth studying at present. In this paper, there are some supervised learning algorithms, but the learning efficiency or applicability is not good enough, in order to improve the learning efficiency, accuracy and ability of the pulse neural network. In order to adapt to more complex problems, this paper combines ReSuMe algorithm, SpikeProp algorithm and other classical algorithms or rules to optimize and innovate the multi-layer network supervised learning algorithm, and the simulation and experiment of the algorithm are carried out. The work of this thesis is as follows: 1) first, the performance of some existing supervised learning algorithms is analyzed, and the accuracy of the algorithm is analyzed. The advantages and disadvantages of degree and other aspects, such as the SpikeProp algorithm, the ReSuMe algorithm and other.2), propose a multi pulse multilayer neural network supervised learning algorithm. This algorithm combines the ReSuMe algorithm to optimize and innovate the existing multi-layer algorithms. Finally, a simulation.3 is carried out on the algorithm. On the basis of this, a neural network monitoring system based on the delay is proposed. The algorithm makes the learning process of the neural network no longer only limited to the adjustment of the weight of the synapse, extends and innovating the original algorithm, and finally simulated the algorithm.4). Finally, the algorithm has been successfully applied to the XOR logic or Iris data collection class and so on, showing good results.
【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】：TP18

【参考文献】