Liquid computing of spiking neural network with multi-clustered and active-neuron-dominant structure
Liquid computing of spiking neural network with multi-clustered and active-neuron-dominant structure
复制标题
多簇主动神经元主导结构的尖峰神经网络的液体计算
DOI：
10.1016/j.neucom.2017.03.022
复制
发表时间：
2017-06-21
期刊：
影响因子：
6
通讯作者：
Song, Yongduan
中科院分区：
文献类型：
--
作者：
Li, Xiumin;Liu, Hui;Song, Yongduan
液态计算是神经网络智能计算的一种有效方法，特别是对于脉冲神经网络。如果液体网络嵌入适当的结构，它可以执行复杂的计算任务。然而，具有更多生物特征的自组织神经网络的建模仍然是一个重要的开放性挑战，导致在提高模型的计算能力和动态多样性方面存在重大限制。在这里，我们提出了一种新的类型的液体计算模型与多簇和激活神经元主导结构的脉冲神经网络，而不是传统的随机结构。研究了聚类生成方法中聚类数和时间窗口大小的最优参数设置。每个簇中的突触权重通过尖峰定时依赖的可塑性规则进一步细化，以获得活跃的神经元主导的结构。结果表明，该模型在计算流体流动时比随机模型有更好的性能。该算法分别从基于聚类结构和基于活动神经元主导结构的活动同步性和网络灵敏度两个方面来提高网络的信息处理能力。统计分析表明，我们提出的网络的结构熵和活动熵都增加了，这表明它具有较高的拓扑复杂性和动力学多样性。因此，确认了该网络的信号传输效率的提高。(C)2017爱思唯尔B.V.保留所有权利。
Liquid computing is an effective approach to intelligent computations of neural networks, especially for spiking neural networks. If the liquid network is embedded with a proper structure it can perform complex computational tasks. However, the modeling of self-organized neural networks with more biological characteristics is still an important open challenge, resulting in major constraints on improving the computational capability and dynamical diversity of the model. Here, we present a novel type of liquid computing model with both multi-clustered and active-neuron-dominant structure of spiking neural network, instead of the traditional random structure. The optimal parameter settings of the cluster number and time window size of clustering generation method had been considered. The synaptic weights in each cluster are further refined through the spike-timing-dependent plasticity rule to obtain an active neuron-dominant structure. The results show that this model has much better performance on liquid computing than the random model. The enhancement of information processing capability is achieved by improving two aspects, i.e. the activity synchrony and network sensitivity, based on the clustered structure and active-neuron-dominant structure, respectively. Statistical analysis demonstrates that both structure entropy and activity entropy of our proposed network are increased, indicating its high topological complexity and dynamical diversity. Therefore, the improvement in efficiency of signal transmission of this network is confirmed. (C) 2017 Elsevier B.V. All rights reserved.