【强基固本】脉冲神经网络(SNN)
“强基固本,行稳致远”,科学研究离不开理论基础,人工智能学科更是需要数学、物理、神经科学等基础学科提供有力支撑,为了紧扣时代脉搏,我们推出“强基固本”专栏,讲解AI领域的基础知识,为你的科研学习提供助力,夯实理论基础,提升原始创新能力,敬请关注。
地址:https://www.zhihu.com/people/AI_team-WSF
本文大部分内容来自:脉冲神经网络的五脏六腑,做一下笔记对原始文章增添一下自己的理解。
https://blog.csdn.net/u011853479/article/details/61414913
第一代神经网络:感知机,第二代神经网络:ANN,第三代神经网络:脉冲神经网络。由于DCNN采用基于速率的编码,所以其硬件实现需要消耗更多的‘能量’。SNN中每个神经元最多使用一个脉冲,而大多数神经元根本不放电,导致能量消耗最小。
Spike-YOLO:SNN在目标检测上的首次尝试:
https://zhuanlan.zhihu.com/p/159982953
SNN的构建过程
学习规则
ANN向SNN转换
https://blog.csdn.net/h__ang/article/details/90609793
01
构建脉冲神经元模型
1.1 Hodgkin-Huxley( HH)模型
1.2 Leaky Integrate and Fire(LIF)模型
https://blog.csdn.net/qq_34886403/article/details/75735448
02
神经脉冲序列
2.1 瞬时放电频率
2.2 齐次泊松过程
2.2.1 泊松脉冲序列的生成
2.2.1.1 相邻脉冲间的时间间隔固定
2.2.1.2 相邻脉冲见的时间间隔不固定
2.3 非齐次泊松过程
2.4 扩展
2.4.1 不应期
2.4.2 涌现
03
脉冲神经网络的训练方法
突触的权重的调整是由突触前和突触后神经元的相对峰值时间决定的以及teacher的反馈。 STDP 分为LTP(长期增强作用)和LTD(长期抑制作用)。 在训练中,无论何时z神经元释放一个峰值,它都会经历某种形式的STDP。 teacher决定输出的z神经元是进行Hebbian STDP还是anti-Hebbian STDP,如果z神经元表示目标类别,那么它将进行Hebbian STDP。否则,它将进行anti-Hebbian STDP。 使用正则化技术对权重进行了重新规范化(最后的改进将会讲到为什么要这样做)。
3.1 无监督学习算法
3.1.1 三相STDP
3.1.2 二相STDP
3.2 有监督学习算法
3.2.1 基于突触可塑性的监督学习算法
3.2.1.1 监督Hebbian学习算法
3.2.1.2 远程监督学习算法(ReSuMe)
3.2.1.3 其他监督学习算法
04
ANN向SNN的转化
4.1 CNN转化为SNN
CNN中的某一层神经元可能会输出负值,这些负数用SNN表示会比较困难。比如,CNN中的sigmoid函数的输出范围为-1到1,每个卷积层的加权和可能也为负值,在进行颜色空间变换时某个颜色通道也可能取到负值。虽然,SNN可以采用抑制神经元来表示负值,但会成倍地增加所需要的神经元数目,提高计算的复杂度。 对于SNN,CNN中神经元的偏置很难表示。 CNN中的Max-pooling层对应到SNN中,需要两层的脉冲神经网络。同样提高了SNN的复杂度。
保证CNN中的每个神经元的输出值都是正数。在图像预处理层之后,卷积层之前加入abs()绝对值函数,保证卷积层的输入值都是非负的。将神经元的激活函数替换为ReLU函数。一方面可以加快原始CNN的收敛速度,另一方面ReLU函数和LIF神经元的性质比较接近,可以最小化网络转化之后的精度损失。 将CNN所有的偏置项都置为0。 采用空间线性降采样层代替Max-pooling层。
SNN的网络结构与裁剪后的CNN相同,单纯地将CNN中的人工神经元替换为脉冲神经元。 在CNN的卷积层之前增加用于脉冲序列生成的一层网络,对输入的图像数据进行编码,转化为脉冲序列。 变换决策方式。在某一段时间内,对全连接层输出的脉冲数进行统计,取脉冲数最多的类别作为最终的分类结果。 将裁减后的CNN训练得到的权值全部迁移到对应的SNN。
神经元的输入脉冲不够,导致神经元的膜电压无法超过设定的阈值,造成放电频率过低。 神经元的输入脉冲过多,导致ReLU模型在每个采样周期内输出多个脉冲。 因为脉冲序列输入是以一定的概率选择的,会导致某些特征一直处于边缘地带或占据了过多的主导地位。
参考文献
[1] Gerstner W, Kistler W M. Spiking neuron models: Single neurons, populations, plasticity[M]. Cambridge university press, 2002.
[2] Izhikevich E M. Simple model of spiking neurons[J]. IEEE Transactions on neural networks, 2003, 14(6): 1569-1572.
[3] Izhikevich E M. Hybrid spiking models[J]. Philosophical Transactions of the Royal Society of London A: Mathematical, Physical and Engineering Sciences, 2010, 368(1930): 5061-5070.
[4] Heeger D. Poisson model of spike generation[J]. Handout, University of Standford, 2000, 5: 1-13.
[5] Hebb D O. The organization of behavior: A neuropsychological theory[M]. Psychology Press, 2005.
[6] Caporale N, Dan Y. Spike timing-dependent plasticity: a Hebbian learning rule[J]. Annu. Rev. Neurosci., 2008, 31: 25-46.
[7] Ponulak F, Kasinski A. Supervised learning in spiking neural networks with ReSuMe: sequence learning, classification, and spike shifting[J]. Neural Computation, 2010, 22(2): 467-510.
[8] Wade J J, McDaid L J, Santos J A, et al. SWAT: a spiking neural network training algorithm for classification problems[J]. IEEE Transactions on neural networks, 2010, 21(11): 1817-1830.
[9] Rumelhart D E, Hinton G E, Williams R J. Learning representations by back-propagating errors[J]. Cognitive modeling, 1988, 5(3): 1.
[10] Bohte S M, Kok J N, La Poutre H. Error-backpropagation in temporally encoded networks of spiking neurons[J]. Neurocomputing, 2002, 48(1): 17-37.
[11] Ghosh-Dastidar S, Adeli H. A new supervised learning algorithm for multiple spiking neural networks with application in epilepsy and seizure detection[J]. Neural networks, 2009, 22(10): 1419-1431.
[12] Mohemmed A, Schliebs S, Matsuda S, et al. Span: Spike pattern association neuron for learning spatio-temporal spike patterns[J]. International Journal of Neural Systems, 2012, 22(04): 1250012.
[13] Mohemmed A, Schliebs S, Matsuda S, et al. Training spiking neural networks to associate spatio-temporal input–output spike patterns[J]. Neurocomputing, 2013, 107: 3-10.
[14] Yu Q, Tang H, Tan K C, et al. Precise-spike-driven synaptic plasticity: Learning hetero-association of spatiotemporal spike patterns[J]. Plos one, 2013, 8(11): e78318.
[15] Cao Y, Chen Y, Khosla D. Spiking deep convolutional neural networks for energy-efficient object recognition[J]. International Journal of Computer Vision, 2015, 113(1): 54-66.
[16] Diehl P U, Neil D, Binas J, et al. Fast-classifying, high-accuracy spiking deep networks through weight and threshold balancing[C]//2015 International Joint Conference on Neural Networks (IJCNN). IEEE, 2015: 1-8.
[17] O’Connor P, Neil D, Liu S C, et al. Real-time classification and sensor fusion with a spiking deep belief network[J]. Neuromorphic Engineering Systems and Applications, 2015: 61.
[18] Maass W, Natschläger T, Markram H. Real-time computing without stable states: A new framework for neural computation based on perturbations[J]. Neural computation, 2002, 14(11): 2531-2560.
[19] LIN X, WANG X, ZHANG N, et al. Supervised Learning Algorithms for Spiking Neural Networks: A Review[J]. Acta Electronica Sinica, 2015, 3: 024.
[20] 顾宗华, 潘纲. 神经拟态的类脑计算研究[J]. 中国计算机学会通讯, 2015, 11(10): 10-20.
[1] Gerstner W, Kistler W M. Spiking neuron models: Single neurons, populations, plasticity[M]. Cambridge university press, 2002.
[2] Izhikevich E M. Simple model of spiking neurons[J]. IEEE Transactions on neural networks, 2003, 14(6): 1569-1572.
[3] Izhikevich E M. Hybrid spiking models[J]. Philosophical Transactions of the Royal Society of London A: Mathematical, Physical and Engineering Sciences, 2010, 368(1930): 5061-5070.
[4] Heeger D. Poisson model of spike generation[J]. Handout, University of Standford, 2000, 5: 1-13.
[5] Hebb D O. The organization of behavior: A neuropsychological theory[M]. Psychology Press, 2005.
[6] Caporale N, Dan Y. Spike timing-dependent plasticity: a Hebbian learning rule[J]. Annu. Rev. Neurosci., 2008, 31: 25-46.
[7] Ponulak F, Kasinski A. Supervised learning in spiking neural networks with ReSuMe: sequence learning, classification, and spike shifting[J]. Neural Computation, 2010, 22(2): 467-510.
[8] Wade J J, McDaid L J, Santos J A, et al. SWAT: a spiking neural network training algorithm for classification problems[J]. IEEE Transactions on neural networks, 2010, 21(11): 1817-1830.
[9] Rumelhart D E, Hinton G E, Williams R J. Learning representations by back-propagating errors[J]. Cognitive modeling, 1988, 5(3): 1.
[10] Bohte S M, Kok J N, La Poutre H. Error-backpropagation in temporally encoded networks of spiking neurons[J]. Neurocomputing, 2002, 48(1): 17-37.
[11] Ghosh-Dastidar S, Adeli H. A new supervised learning algorithm for multiple spiking neural networks with application in epilepsy and seizure detection[J]. Neural networks, 2009, 22(10): 1419-1431.
[12] Mohemmed A, Schliebs S, Matsuda S, et al. Span: Spike pattern association neuron for learning spatio-temporal spike patterns[J]. International Journal of Neural Systems, 2012, 22(04): 1250012.
[13] Mohemmed A, Schliebs S, Matsuda S, et al. Training spiking neural networks to associate spatio-temporal input–output spike patterns[J]. Neurocomputing, 2013, 107: 3-10.
[14] Yu Q, Tang H, Tan K C, et al. Precise-spike-driven synaptic plasticity: Learning hetero-association of spatiotemporal spike patterns[J]. Plos one, 2013, 8(11): e78318.
[15] Cao Y, Chen Y, Khosla D. Spiking deep convolutional neural networks for energy-efficient object recognition[J]. International Journal of Computer Vision, 2015, 113(1): 54-66.
[16] Diehl P U, Neil D, Binas J, et al. Fast-classifying, high-accuracy spiking deep networks through weight and threshold balancing[C]//2015 International Joint Conference on Neural Networks (IJCNN). IEEE, 2015: 1-8.
[17] O’Connor P, Neil D, Liu S C, et al. Real-time classification and sensor fusion with a spiking deep belief network[J]. Neuromorphic Engineering Systems and Applications, 2015: 61.
[18] Maass W, Natschläger T, Markram H. Real-time computing without stable states: A new framework for neural computation based on perturbations[J]. Neural computation, 2002, 14(11): 2531-2560.
[19] LIN X, WANG X, ZHANG N, et al. Supervised Learning Algorithms for Spiking Neural Networks: A Review[J]. Acta Electronica Sinica, 2015, 3: 024.
[20] 顾宗华, 潘纲. 神经拟态的类脑计算研究[J]. 中国计算机学会通讯, 2015, 11(10): 10-20.
本文目的在于学术交流,并不代表本公众号赞同其观点或对其内容真实性负责,版权归原作者所有,如有侵权请告知删除。
直播预告
“强基固本”历史文章
分享、点赞、在看,给个三连击呗!