【作者】 王宁；

【导师】 杨洁；

【作者基本信息】 大连理工大学 ， 计算数学， 2019， 硕士

【摘要】 人工神经网络(ANNs)中的神经元特征在于单个静态的连续值激活。然而,生物神经元使用离散脉冲尖峰来计算和传输信息,除了脉冲速率之外,脉冲时间也很重要。因此,脉冲神经网络(SNNs)比人工神经网络在生物学上更具现实性,如果想要了解大脑如何在神经元描述水平上进行计算,可以说它是唯一可行的选择。生物神经元的脉冲在时间和空间上都是稀疏的,并且是事件驱动的。结合生物合理的局部学习规则,这使得为脉冲神经网络构建低功耗神经形态硬件变得更加容易。脉冲神经元的传递函数通常是不可微分的,这阻止了使用反向传播。近年来,SNNs学习算法也在被不断地探索。SNNs中的无监督学习通常将STDP作为其学习机制的一部分。最常见的生物STDP形式具有非常直观的解释。如果突触前神经元在突触后神经元之前短暂地发射,则连接它们的权重会增强。如果突触前神经元在突触后神经元后短暂发射,则时间事件之间的因果关系是虚假的,并且权重减弱。大多数现有的监督学习算法都是基于具有固有缺陷,例如局部最优和过度拟合,的梯度下降学习算法。于是,在脉冲网络上寻找具有全局最优解的学习算法成为当前研究热门问题。在本文中,我们研究了涉及模拟退火机制的遗传算法作为SNN的监督训练算法的性能。本算法关键的思想在于采用全局搜索,有效避免局部最优和过度拟合,同时避免反向梯度的计算,没有主动引入误差。我们对网络训练过程和网络参数进行了细致分析,找到了最优参数组合。根据实验结果,该方法在众所周知的分类问题上具有比其他学习算法更高的准确性。更多还原

【Abstract】 Neurons in an Artificial neural networks(ANNs)are characterized by a single,static,continuous valued activation.Yet biological neurons use discrete spikes to compute and transmit information,and the spike times,in addition to the spike rates,matter.Spiking neural networks(SNNs)are thus more biologically realistic than ANNs,and are arguably the only viable option if one wants to understand how the brain computes at the neuronal description level.The spikes of biological neurons are sparse in time and space,and event-driven.Combined with bio-plausible local learning rules,this makes it easier to build low-power,neuromorphic hardware for SNNs.Spiking neurons’ transfer function is usually non-differentiable,which prevents using back-propagation.In recent years,spiking neural network learning algorithms are also being continuously explored.Unsupervised learning in SNNs often involves spike-timing-dependent-plasticity(STDP)as part of the learning mechanism.The most common form of biological STDP has a very intuitive interpretation.If a presynaptic neuron fires briefly before the postsynaptic neuron,the weight connecting them is strengthened.If the presynaptic neuron fires briefly after the postsynaptic neuron,then the causal relationship between the temporal events is spurious and the weight is weakened.Most existing training supervised algorithms are based on gradient descent with inherent defects,such as local optimum and over-fitting.Therefore,finding a learning algorithm with a global optimal solution has become a hot topic in current research.In this paper,we investigate the performance of the Genetic Algorithm Involving Mechanism of Simulated Annealing,as a supervised training algorithm for SNNs.The key idea is to adopt global search,which effectively avoid local optima and over-fitting.At the same time,we conducted a detailed analysis of the network training process and network parameters,and found the optimal combination of parameters.According to the experiment results,this approach has higher accuracy than other learning algorithms on well-known classification problems.更多还原