一种非局部连接的抗噪性储备池构建方法

Construction of a non-locally connected anti-noise reservoir pool

中文关键词  反馈神经网络  储备池计算  脉冲神经网络  生物真实性  液体状态机  小样本学习

英文关键词  recurrent neural networks  reservoir computing  spiking neural network  biological plausible  liquid state machine  limited sample learning

中文摘要

      由于具有高度的生物真实性, 液体状态机在抗噪性、鲁棒性方面相对于人工神经网络具有更大的优势, 但 也更难优化. 采用人工神经网络思想对液体状态机进行的优化, 牺牲了生物真实性和网络泛用性的同时, 并不能保 证优化的有效性; 而依据生物神经系统内抽象出的规律进行储备池的优化, 则优化算法非常复杂. 为了提高储备池 的泛用性和抗噪性, 同时避免复杂的优化过程, 本文模拟大脑中普遍存在的各神经元集群间的非局部连接分布—伽 马分布来生成储备池的权值, 生成一个具有更高生物真实性、隐含功能柱结构的储备池. 首先, 通过对储备池活动 和储备池进行Lempel-Ziv复杂度分析, 从理论上说明该种储备池权值生成方式的优势; 然后, 通过与脉冲时序可塑 性算法(STDP)和高斯分布等进行对比实验, 证明本文采用伽马分布生成的储备池具有更高的准确度和更强的抗噪 性.

英文摘要

      Due to its high bio-plausibility, the liquid state machine (LSM) has greater advantages in terms of noise immunity and robustness than artificial neural networks (ANN), but it is also more difficult to optimize. On one hand, if the liquid state machine is optimized in the same way as optimizing ANN, the bio-plausibility of the reservoir pool and the generalization of the network are sacrificed, and the effectiveness of the optimization cannot be guaranteed; on the other hand, if the reservoir pool is optimized according to rules that extracted from the biological neural system, the optimizing algorithm will be very complicated. In order to improve the generalization and noise immunity of the reservoir pool while avoiding the complicated optimization, the gamma distribution, a common non-local connection distribution between the various neuron clusters in the brain, is used to generate the weight of the reservoir pool, and a reservoir pool with higher bio-plausibility and implicit functional column structure is generated. Firstly, by analysing the Lempel-Ziv complexity of the reservoir pool and its activity, the advantages of the proposed weight generation method are theoretically explained. Then, compared with the experimental results of spiking-timing dependent plasticity (STDP) and Gaussian distribution, it is proved that reservoir pool generated according to the gamma distribution has higher accuracy and stronger noise immunity.