脉冲神经网络的新结构与学习算法研究

With the high simulation to biological neural systems, the spiking neural network has increasingly attracted attentions and is called the third generation neural network. However, it has a large computation and its accuracy is easily affected by the network's structure. These disadvantages are considered the brief bottlenecks of this network. The aim of this project is to improve the network and to accelerate learning through the studies of network constitution methods and learning algorithms. Firstly, the problem of embedding the spike stimulating strength (the gradient of the state function at the firing time) is studied. With this approach, we will explore the influence degree and pattern of the spike stimulating strength on the information processing ability of the postsynaptic neuron. Meanwhile, we will further seek other factors which effect the neuron stimulating to present improved network operation mechanism. Secondly, with the research results funded by the prior National Natural Science Foundation, we will study the fuzzification of spiking neural networks and present several network structures which are based on s-t norms to expand the information processing ability for spiking neural networks. In addition, as for learning algorithms, we have proved that the traditional BP algorithm is fully suitable for the spiking neural network without any assumption. Based on this result, we hope to analyze the BP algorithm and other learning algorithms and to explore the possibilities of more efficient learning algorithms in the spiking neural network.

脉冲神经网络由于对生物神经系统的高度仿真性已经受到越来越多的关注，被称为第三代神经网络。本项目旨在研究新的脉冲神经网络结构和更好的学习算法。首先，针对我们最近提出的一种新的网络结构,研究如何将脉冲激发强度(即状态函数x(t)达到阈值时的导数)嵌入到脉冲网络的网络构造与运行机制中去，探索脉冲激发强度对突触后神经元信息处理能力的影响程度与影响方式，同时进一步探寻其它影响脉冲激发因素，使脉冲神经网络在不影响训练精度前提下减小网络规模,改善推广精度。其次，结合申请者主持的前一个国家自然科学基金的工作,研究脉冲神经网络的模糊化问题，提出几种基于s-t模的模糊脉冲神经网络结构,从而扩展脉冲神经网络的信息处理能力。另外，我们已经证明传统BP算法也完全适用于脉冲神经网络，希望以此为基础,对脉冲神经网络的BP算法及其它学习算法进行分析与比较,探讨适用于脉冲神经网络的更高效的学习算法。

脉冲神经网络是一种对生物神经系统高度仿真的网络，目前已经受到越来越多的关注，被称为第三代神经网络。然而脉冲网络在计算过程中需要大量的计算，而且精确性极易受到网络结构的影响。本项目的研究计划是通过研究网络的性能提出新的脉冲神经网络结构和更好的学习算法。在具体的研究工作中，针对脉冲网络分类能力的鲁棒性，我们目前已经取得了一些成果并发表在国际期刊上。我们通过在脉冲神经网络的输入信号中加入了两种典型的扰动，正弦扰动与高斯扰动，来研究网络分类能力的鲁棒性，这方面的研究在当前的脉冲网络研究中比较少。一系列数值试验表明，在脉冲神经网络的输入信号受到一定干扰的情况下，它的分类能力不会发生大的变化，但是对干扰具有一定的敏感性。另外，我们提出了基于极端学习机和L_1/2正则化的学习方法来修剪双并联神经网络的隐层节点数目，希望能将这一结果应用到脉冲网络中。在本项目的资助下我们的第三项工作是提出了用蚁群算法以提高测地线的精度，这一结果也已发表为论文。