搜索

x

留言板

尊敬的读者、作者、审稿人, 关于本刊的投稿、审稿、编辑和出版的任何问题, 您可以本页添加留言。我们将尽快给您答复。谢谢您的支持!

姓名
邮箱
手机号码
标题
留言内容
验证码

神经网络超参数优化的删除垃圾神经元策略

黄颖 顾长贵 杨会杰

引用本文:
Citation:

神经网络超参数优化的删除垃圾神经元策略

黄颖, 顾长贵, 杨会杰

Junk-neuron-deletion strategy for hyperparameter optimization of neural networks

Huang Ying, Gu Chang-Gui, Yang Hui-Jie
PDF
HTML
导出引用
  • 随着深度学习处理问题的日益复杂, 神经网络的层数、神经元个数、和神经元之间的连接逐渐增加, 参数规模急剧膨胀, 优化超参数来提高神经网络的预测性能成为一个重要的任务. 文献中寻找最优参数的方法如灵敏度剪枝、网格搜索等, 算法复杂而且计算量庞大. 本文提出一种超参数优化的“删除垃圾神经元策略”. 权重矩阵中权重均值小的神经元, 在预测中的贡献可以忽略, 称为垃圾神经元. 该策略就是通过删除这些垃圾神经元得到精简的网络结构, 来有效缩短计算时间, 同时提高预测准确率和模型泛化能力. 采用这一策略, 长短期记忆网络模型对几种典型混沌动力系统的预测性能得到显著改善.
    With the complexity of problems in reality increasing, the sizes of deep learning neural networks, including the number of layers, neurons, and connections, are increasing in an explosive way. Optimizing hyperparameters to improve the prediction performance of neural networks has become an important task. In literatures, the methods of finding optimal parameters, such as sensitivity pruning and grid search, are complicated and cost a large amount of computation time. In this paper, a hyperparameter optimization strategy called junk neuron deletion is proposed. A neuron with small mean weight in the weight matrix can be ignored in the prediction, and is defined subsequently as a junk neuron. This strategy is to obtain a simplified network structure by deleting the junk neurons, to effectively shorten the computation time and improve the prediction accuracy and model the generalization capability. The LSTM model is used to train the time series data generated by Logistic, Henon and Rossler dynamical systems, and the relatively optimal parameter combination is obtained by grid search with a certain step length. The partial weight matrix that can influence the model output is extracted under this parameter combination, and the neurons with smaller mean weights are eliminated with different thresholds. It is found that using the weighted mean value of 0.1 as the threshold, the identification and deletion of junk neurons can significantly improve the prediction efficiency. Increasing the threshold accuracy will gradually fall back to the initial level, but with the same prediction effect, more operating costs will be saved. Further reduction will result in prediction ability lower than the initial level due to lack of fitting. Using this strategy, the prediction performance of LSTM model for several typical chaotic dynamical systems is improved significantly.
      通信作者: 杨会杰, hjyang@usst.edu.cn
    • 基金项目: 国家自然科学基金(批准号: 11875042, 11505114)资助的课题.
      Corresponding author: Yang Hui-Jie, hjyang@usst.edu.cn
    • Funds: Project supported by the National Natural Science Foundation of China (Grant Nos. 11875042, 11505114).
    [1]

    邓帅 2019 计算机应用研究 36 1984

    Deng S 2019 Appl. Res. Comput. 36 1984

    [2]

    邵恩泽, 吴正勇, 王灿 2020 工业控制计算机 33 11Google Scholar

    Shao E Z, Wu Z Y, Wang C 2020 Ind. Contrl. Comput. 33 11Google Scholar

    [3]

    乔俊飞, 樊瑞元, 韩红桂, 阮晓钢 2010 控制理论与应用 27 111

    Qiao J F, Fan R Y, Han H G, Ruan X G 2010 Contl. Theor. Appl. 27 111

    [4]

    陈国茗, 于腾腾, 刘新为 2021 数值计算与计算机应用 42 215Google Scholar

    Chen G M, Yu T T, Liu X W 2021 J. Num. Method. Comp. Appl. 42 215Google Scholar

    [5]

    魏德志, 陈福集, 郑小雪 2015 物理学报 64 110503Google Scholar

    Wei D Z, Chen F J, Zheng X X 2015 Acta Phys. Sin. 64 110503Google Scholar

    [6]

    王新迎, 韩敏 2015 物理学报 64 070504Google Scholar

    Wang X Y, Han M 2015 Acta Phys. Sin 64 070504Google Scholar

    [7]

    黄伟建, 李永涛, 黄远 2021 物理学报 70 010501Google Scholar

    Huang W J, Li Y T, Huang Y 2021 Acta Phys. Sin. 70 010501Google Scholar

    [8]

    Yamaguti Y, Tsuda I 2021 Chaos 31 013137Google Scholar

    [9]

    Graves A 2013 arXiv: 1308.0850 [cs. NE]

    [10]

    Johnston D E 1978 Proc 8 th BHRA Int Conf Fluid Sealing Durham, UK, 1978 pC1-1

    [11]

    Sezer O B, Gudelek M U, Ozbayoglu A M 2020 Appl. Soft Comput. J. 90 106181Google Scholar

    [12]

    甘文娟, 陈永红, 韩静, 王亚飞 2020 计算机系统应用 29 212Google Scholar

    Gan W J, Chen Y H, Han J, Wang Y F 2020 Comput. Syst. Appl. 29 212Google Scholar

    [13]

    Farmelo G 2002 It Must Be Beautiful: Great Equations of Modern Science (London: Granta Publications) pp28–45

    [14]

    Grassberger P, Procaccia I 1983 Physica D 9 189Google Scholar

    [15]

    Nauenberg M 1983 Ann. N. Y. Acad. Sci. 410 317Google Scholar

    [16]

    张中华, 丁华福 2009 计算机技术与发展 19 185Google Scholar

    Zhang Z H, Ding H F 2009 Comput. Technol. Dev. 19 185Google Scholar

    [17]

    Butcher J C 1967 J. ACM 14 84Google Scholar

    [18]

    Liu C, Yin S Q, Zhang M, Zeng Y, Liu J Y 2014 Appl. Mech. Mater. 644-650 2216

    [19]

    Bao Y K, Liu Z T 2006 LNCS 4224 504

    [20]

    Ou Y Y, Chen G H, Oyang Y J 2006 LNCS 4099 1017

  • 图 1  LSTM神经网络 (a) LSTM模型网络结构; (b)单元内部运行逻辑

    Fig. 1.  LSTM neural network: (a) network structure of LSTM; (b) run logic inside the cell.

    图 2  混沌时间序列 (a) Logistic系统, μ$ =3.9 $; (b) Henon系统; (c) Rossler系统

    Fig. 2.  Chaotic time series: (a) Logistic system, μ$ =3.9 $; (b) Henon system; (c) Rossler system

    图 3  隐藏层输入神经元对输出神经元的权重热图

    Fig. 3.  Heat map of weight of input neuron to output neuron in hidden layer.

    图 4  隐藏层各输入神经元对输出神经元的权重均值热图

    Fig. 4.  Heat map of weights’ mean value of input neurons to output neurons in hidden layer.

    图 5  权重均值折线图

    Fig. 5.  Line graph of the weights’ mean.

    图 6  不同混沌状态对应的最优权重阈值变化

    Fig. 6.  The change of optimal weight threshold corresponding to different chaotic states.

    图 7  不同混沌状态对应的预测准确率变化

    Fig. 7.  The change of prediction accuracy of different chaotic states.

    表 1  模型参数及结果

    Table 1.  Parameters and results of the models.

    模型 train test win L1 L2 D1 D2 准确率
    Logistic
    (μ = 3.6)
    5000 15 22 16 4 4 1 82.90%
    Logistic
    (μ = 3.7)
    5000 15 22 16 4 4 1 70.70%
    Logistic
    (μ = 3.8)
    500015 2 20 4 4 1 68.60%
    Logistic
    (μ = 3.9)
    500015 2 16 4 2 1 60.00%
    Logistic
    (μ = 3.99)
    500015 2 16 4 4 1 57.10%
    Henon500015 2 22 2 2 1 67.10%
    Rossler500015 2 16 4 4 1 77.10%
    下载: 导出CSV

    表 2  权重结构拆分

    Table 2.  Weight structure resolution.

    结构起始点终点
    input_gate0units
    forget_gateunits2×units
    cell2×units3×units
    output_gate3×units4×units
    下载: 导出CSV

    表 3  输出门权重矩阵图

    Table 3.  Heat diagram of output door’s weights.

    w1 w2 w3 w4 w5 w6 w7 w8 w9 w10 w11 w12 w13 w14 w15 w16
    x1 0.032129 0.144328 0.007293 0.003557 0.070465 0.011652 0.11142 0.048287 0.105727 0.039649 0.010162 0.119949 0.0771 0.126937 0.001947 0.089396
    x2 0.070867 0.013232 0.054503 0.073496 0.005734 0.064071 0.145835 0.083217 0.025553 0.033065 0.113584 0.028327 0.054085 0.177761 0.010123 0.046546
    x3 0.083647 0.046366 0.066226 0.044534 0.010054 0.077919 0.071001 0.018548 0.035294 0.071642 0.053888 0.028448 0.102273 0.096744 0.078535 0.114879
    x4 0.039123 0.113002 0.161581 0.113866 0.070651 0.085571 0.01836 0.015408 0.032978 0.018375 0.068342 0.059137 0.00701 0.070451 0.032602 0.099853
    x5 0.013762 0.055823 0.147481 0.055493 0.093444 0.027264 0.082384 0.058674 0.032903 0.033342 0.045937 0.035937 0.110378 0.004487 0.161653 0.041037
    x6 0.038079 0.089649 0.013102 0.021696 0.042833 0.109787 0.001024 0.0673 0.036916 0.134038 0.066291 0.146953 0.009803 0.081372 0.098701 0.042456
    x7 0.013865 0.001702 0.057869 0.072264 0.104456 0.029761 0.062669 0.07539 0.05715 0.102282 0.017876 0.012626 0.020022 0.100243 0.153076 0.11875
    x8 0.002042 0.006925 0.060008 0.13031 0.136406 0.056203 0.061606 0.028064 0.006926 0.099129 0.055122 0.117276 0.06846 0.014505 0.078184 0.078834
    x9 0.044205 0.171724 0.153162 0.13818 0.029189 0.025947 0.049391 0.012338 0.100584 0.028133 0.004946 0.017914 0.008463 0.014741 0.066944 0.134139
    x10 0.094619 0.0563 0.040223 0.096283 0.10152 0.145036 0.051991 0.075623 0.075216 0.061209 0.057986 0.066076 0.004787 0.01945 0.042341 0.011339
    x11 0.078909 0.005603 0.08149 0.125202 0.069081 0.10143 0.122451 0.027058 0.057647 0.016226 0.03275 0.050667 0.036795 0.11072 0.081767 0.002204
    x12 0.046159 0.005688 0.006237 0.004618 0.014815 0.005272 0.04598 0.037005 0.190933 0.065535 0.005131 0.015155 0.065812 0.099804 0.172294 0.21956
    x13 0.111226 0.027026 0.027497 0.074868 0.139154 0.084413 0.080342 0.038769 0.088824 0.047083 0.056548 0.002081 0.10549 0.049929 0.020529 0.04622
    x14 0.092772 0.03999 0.055938 0.128114 0.036386 0.013061 0.083943 0.051033 0.106374 0.007257 0.063049 0.091929 0.084821 0.020458 0.089496 0.035379
    x15 0.131586 0.038161 0.176303 0.041758 0.049173 0.096633 0.0033 0.045529 0.084262 0.050839 0.003322 0.063406 0.029601 0.045323 0.116047 0.024757
    x16 0.070289 0.05378 0.092472 0.03372 0.052087 0.110236 0.055639 0.124221 0.029371 0.06142 0.04904 0.043376 0.012261 0.041226 0.109564 0.061299

    0.060205 0.054331 0.075087 0.072372 0.064091 0.065266 0.065459 0.050404 0.066666 0.054326 0.043998 0.056204 0.049428 0.067134 0.082113 0.072915
    下载: 导出CSV

    表 4  μ = 3.99时不同参数的预测准确率

    Table 4.  The prediction accuracy of different parameters when μ = 3.99.

    指标 调整前以各权重阀值调整后变化趋势
    0.090.10.11
    L116151210
    准确率57.10%59.30%56.40%51.40%
    下载: 导出CSV

    表 5  神经元数及预测准确率变化表

    Table 5.  Table of neuron numbers and prediction accuracy.

    模型 调整前L1 调整后L1 调整前准确率 调整后准确率 神经元数调整准确率变化趋势
    Logistic
    (μ = 3.6)
    16 15 82.90% 90.70% –1
    Logistic
    (μ = 3.7)
    16 13 70.70% 71.40% –3
    Logistic
    (μ = 3.8)
    2016 68.60% 68.60% –4
    Logistic
    (μ = 3.9)
    1612 60.00% 60.00% –4
    Logistic
    (μ = 3.99)
    1615 57.10% 59.30% –1
    Henon 22 21 67.10% 70.00% –1
    Rossler1614 77.10% 83.60% –2
    下载: 导出CSV

    表 6  μ = 3.6 时不同参数的预测准确率

    Table 6.  The prediction accuracy of different parameters when μ = 3.6.

    指标 调整前以各权重阀值调整后变化趋势
    0.080.090.095
    L116151210
    准确率82.90%90.70%87.90%78.60%
    下载: 导出CSV

    表 7  μ = 3.7 时不同参数的预测准确率

    Table 7.  The prediction accuracy of different parameters when μ = 3.7.

    指标 调整前以各权重阀值调整后变化趋势
    0.0750.090.105
    L11613119
    准确率70.70%71.40%65.00%60.70%
    下载: 导出CSV

    表 8  μ = 3.8时不同参数的预测准确率

    Table 8.  The prediction accuracy of different parameters when μ = 3.8.

    指标 调整前以各权重阀值调整后变化趋势
    0.0850.0950.105
    L120181612
    准确率68.60%68.60%68.60%65.00%
    下载: 导出CSV

    表 9  μ = 3.9 时不同参数的预测准确率

    Table 9.  The prediction accuracy of different parameters when μ = 3.9.

    指标 调整前以各权重阀值调整后变化趋势
    0.090.0950.1
    L116141210
    准确率60.00%60.00%60.00%55.00%
    下载: 导出CSV

    表 10  Henon系统取不同参数的预测准确率

    Table 10.  Prediction accuracy of Henon system for different parameters.

    指标 调整前以各权重阀值调整后变化趋势
    0.10.110.12 0.14
    L122 21191614
    准确率67.10% 70.00%66.40%65.70%65.00%
    下载: 导出CSV

    表 11  Rossler系统取不同参数的预测准确率

    Table 11.  Prediction accuracy of Rossler system for different parameters.

    指标调整前以各权重阀值调整后变化趋势
    0.0850.0950.105 0.115
    L116 1412118
    准确率77.10% 83.60%81.40%80.70%71.40%
    下载: 导出CSV
  • [1]

    邓帅 2019 计算机应用研究 36 1984

    Deng S 2019 Appl. Res. Comput. 36 1984

    [2]

    邵恩泽, 吴正勇, 王灿 2020 工业控制计算机 33 11Google Scholar

    Shao E Z, Wu Z Y, Wang C 2020 Ind. Contrl. Comput. 33 11Google Scholar

    [3]

    乔俊飞, 樊瑞元, 韩红桂, 阮晓钢 2010 控制理论与应用 27 111

    Qiao J F, Fan R Y, Han H G, Ruan X G 2010 Contl. Theor. Appl. 27 111

    [4]

    陈国茗, 于腾腾, 刘新为 2021 数值计算与计算机应用 42 215Google Scholar

    Chen G M, Yu T T, Liu X W 2021 J. Num. Method. Comp. Appl. 42 215Google Scholar

    [5]

    魏德志, 陈福集, 郑小雪 2015 物理学报 64 110503Google Scholar

    Wei D Z, Chen F J, Zheng X X 2015 Acta Phys. Sin. 64 110503Google Scholar

    [6]

    王新迎, 韩敏 2015 物理学报 64 070504Google Scholar

    Wang X Y, Han M 2015 Acta Phys. Sin 64 070504Google Scholar

    [7]

    黄伟建, 李永涛, 黄远 2021 物理学报 70 010501Google Scholar

    Huang W J, Li Y T, Huang Y 2021 Acta Phys. Sin. 70 010501Google Scholar

    [8]

    Yamaguti Y, Tsuda I 2021 Chaos 31 013137Google Scholar

    [9]

    Graves A 2013 arXiv: 1308.0850 [cs. NE]

    [10]

    Johnston D E 1978 Proc 8 th BHRA Int Conf Fluid Sealing Durham, UK, 1978 pC1-1

    [11]

    Sezer O B, Gudelek M U, Ozbayoglu A M 2020 Appl. Soft Comput. J. 90 106181Google Scholar

    [12]

    甘文娟, 陈永红, 韩静, 王亚飞 2020 计算机系统应用 29 212Google Scholar

    Gan W J, Chen Y H, Han J, Wang Y F 2020 Comput. Syst. Appl. 29 212Google Scholar

    [13]

    Farmelo G 2002 It Must Be Beautiful: Great Equations of Modern Science (London: Granta Publications) pp28–45

    [14]

    Grassberger P, Procaccia I 1983 Physica D 9 189Google Scholar

    [15]

    Nauenberg M 1983 Ann. N. Y. Acad. Sci. 410 317Google Scholar

    [16]

    张中华, 丁华福 2009 计算机技术与发展 19 185Google Scholar

    Zhang Z H, Ding H F 2009 Comput. Technol. Dev. 19 185Google Scholar

    [17]

    Butcher J C 1967 J. ACM 14 84Google Scholar

    [18]

    Liu C, Yin S Q, Zhang M, Zeng Y, Liu J Y 2014 Appl. Mech. Mater. 644-650 2216

    [19]

    Bao Y K, Liu Z T 2006 LNCS 4224 504

    [20]

    Ou Y Y, Chen G H, Oyang Y J 2006 LNCS 4099 1017

  • [1] 齐乐天, 王世元, 沈明琳, 黄刚毅. 基于Nyström柯西核共轭梯度算法的混沌时间序列预测. 物理学报, 2022, 71(10): 108401. doi: 10.7498/aps.71.20212274
    [2] 王世元, 史春芬, 钱国兵, 王万里. 基于分数阶最大相关熵算法的混沌时间序列预测. 物理学报, 2018, 67(1): 018401. doi: 10.7498/aps.67.20171803
    [3] 李瑞国, 张宏立, 范文慧, 王雅. 基于改进教学优化算法的Hermite正交基神经网络混沌时间序列预测. 物理学报, 2015, 64(20): 200506. doi: 10.7498/aps.64.200506
    [4] 唐舟进, 彭涛, 王文博. 一种基于相关分析的局域最小二乘支持向量机小尺度网络流量预测算法. 物理学报, 2014, 63(13): 130504. doi: 10.7498/aps.63.130504
    [5] 唐舟进, 任峰, 彭涛, 王文博. 基于迭代误差补偿的混沌时间序列最小二乘支持向量机预测算法. 物理学报, 2014, 63(5): 050505. doi: 10.7498/aps.63.050505
    [6] 赵永平, 王康康. 具有增加删除机制的正则化极端学习机的混沌时间序列预测. 物理学报, 2013, 62(24): 240509. doi: 10.7498/aps.62.240509
    [7] 王新迎, 韩敏, 王亚楠. 含噪混沌时间序列预测误差分析. 物理学报, 2013, 62(5): 050504. doi: 10.7498/aps.62.050504
    [8] 高光勇, 蒋国平. 采用优化极限学习机的多变量混沌时间序列预测. 物理学报, 2012, 61(4): 040506. doi: 10.7498/aps.61.040506
    [9] 张文专, 龙文, 焦建军. 基于差分进化算法的混沌时间序列预测模型参数组合优化. 物理学报, 2012, 61(22): 220506. doi: 10.7498/aps.61.220506
    [10] 王新迎, 韩敏. 基于极端学习机的多变量混沌时间序列预测. 物理学报, 2012, 61(8): 080507. doi: 10.7498/aps.61.080507
    [11] 宋彤, 李菡. 基于小波回声状态网络的混沌时间序列预测. 物理学报, 2012, 61(8): 080506. doi: 10.7498/aps.61.080506
    [12] 马军, 苏文涛, 高加振. Hindmarsh-Rose混沌神经元自适应同步和参数识别的优化研究. 物理学报, 2010, 59(3): 1554-1561. doi: 10.7498/aps.59.1554
    [13] 张春涛, 马千里, 彭宏. 基于信息熵优化相空间重构参数的混沌时间序列预测. 物理学报, 2010, 59(11): 7623-7629. doi: 10.7498/aps.59.7623
    [14] 宋青松, 冯祖仁, 李人厚. 用于混沌时间序列预测的多簇回响状态网络. 物理学报, 2009, 58(7): 5057-5064. doi: 10.7498/aps.58.5057
    [15] 张勇, 关伟. 基于最大Lyapunov指数的多变量混沌时间序列预测. 物理学报, 2009, 58(2): 756-763. doi: 10.7498/aps.58.756
    [16] 王永生, 孙 瑾, 王昌金, 范洪达. 变参数混沌时间序列的神经网络预测研究. 物理学报, 2008, 57(10): 6120-6131. doi: 10.7498/aps.57.6120
    [17] 韩 敏, 史志伟, 郭 伟. 储备池状态空间重构与混沌时间序列预测. 物理学报, 2007, 56(1): 43-50. doi: 10.7498/aps.56.43
    [18] 李 军, 刘君华. 一种新型广义RBF神经网络在混沌时间序列预测中的研究. 物理学报, 2005, 54(10): 4569-4577. doi: 10.7498/aps.54.4569
    [19] 胡玉霞, 高金峰. 一种预测混沌时间序列的模糊神经网络方法. 物理学报, 2005, 54(11): 5034-5038. doi: 10.7498/aps.54.5034
    [20] 谭 文, 王耀南, 周少武, 刘祖润. 混沌时间序列的模糊神经网络预测. 物理学报, 2003, 52(4): 795-801. doi: 10.7498/aps.52.795
计量
  • 文章访问数:  2411
  • PDF下载量:  54
  • 被引次数: 0
出版历程
  • 收稿日期:  2022-03-10
  • 修回日期:  2022-04-21
  • 上网日期:  2022-08-12
  • 刊出日期:  2022-08-20

/

返回文章
返回