摘要： 远程有监督方法(ReSuMe)通过计算神经元运行时的输出脉冲和输入脉冲的时间差调整突触权值,是目前在理论基础和实际应用上都较出色的脉冲神经元有监督学习方法,但是当期望输出脉冲序列较长时,ReSuMe方法的学习精度较低。为解决该问题,分析影响ReSuMe方
法性能的2个主要因素:在线、离线学习方式及学习过程中更新突触权值时输入脉冲的选取。在线学习精度一般高于离线学习,但是学习精度的差异随着参数或者其他设置的不同有较大差别。针对输入脉冲的选取,提出一种新的学习策略以改进ReSuMe方法,该策略在计算权值
调整幅度时综合考虑期望输出与实际输出脉冲序列,从而避免增强与减弱权值时输入脉冲出现重叠干扰。实验结果表明,新的学习策略可以有效提高ReSuMe方法的学习精度及其解决实际问题的能力。
关键词:
脉冲神经元,
远程有监督方法,
脉冲序列学习,
脉冲神经网络,
脉冲反应模型
Abstract: Remote Supervised Method(ReSuMe) adjusts the synaptic weights according to the time difference between output and input spiking during neuron runtime.Whether for the theoretical basis or the practical application,ReSuMe method is a kind of relatively excellent supervised learning method for spiking neurons. But when the desired output spriking sequence is long,the learning accuracy of ReSuMe method is relatively low. In order to discusse ReSuMe method more detailed and improve its learning performance,two factors that affect the learning performance of ReSuMe method are studied. The first factor is the offline or online learning model,the other is the choice of the input spiking considered when adjusting the synaptic weights. Online learning accuracy is generally higher than offline learning, but the different of learning precision is big with different parameters or other settings. And a new learning strategy is proposed for the second factor to improve the ReSuMe method, it considers input and output spiking sequence when calculates weight adjustment,avoids the overlap interference of input spiking when increases and decreases weights. Experimental results show that the new strategy can greatly improve the learning accuracy of ReSuMe method under different experiment settings. So it has a strong ability to solve practical problems.
Key words:
spiking neuron,
Remote Supervised Method(ReSuMe),
spiking sequence learning,
spiking neural network,
Spiking Response Model(SRM)
中图分类号: