数据–知识融合的机器学习(1):模型分析
|
摘要
数据驱动的机器学习模型往往依赖于大幅提高训练样本数量以降低其泛化风险,对有限样本的应用场景适用性差。为此,该文在数据驱动的模型基础上引入领域知识这一学习要素,提出数据–知识融合的机器学习范式,可以降低机器学习方法的泛化风险。首先,给出关于数据–知识融合机器学习问题的表示,分析数据–知识融合的多种不同模式,并建立一般性的数据–知识融合机器学习模型。然后,分析该融合学习模型的解的形态,给出评价该模型在问题的全域学习空间和局部学习空间的泛化能力的表示。最后,结合实际应用案例,讨论该融合学习模型在回归分析、模式识别、动态规划等任务中如何实现数据–知识的融合学习。与单纯数据驱动的模型相比,数据–知识融合模型可使机器学习过程更加高效,并且在不提高训练样本数量的前提下降低学习器泛化风险。
In machine learning, the main strategy for reducing the generalization risk of the data-driven model(DDM) is to increase the training samples. In this paper, an alternative method to reduce the generalization risk was proposed, in which domain knowledges are embedded into the data-driven learning paradigm and it is particularly useful for the scenarios with limited training samples. First, we introduce the formulation of the data-driven model incorporated with knowledge(KDM), and its solutions. Then, the characteristics of the KDM's solutions were analyzed, and the methods to measure KDM's generalization capability in both the global learning space and local learning space were given. Finally, the applicability of KDM in regression analysis, pattern recognition, and dynamic programming problems were discussed. Compared with DDM, the learning procedure in KDM is more efficient and can reduce the generalization risk without increasing the training samples.
In machine learning, the main strategy for reducing the generalization risk of the data-driven model(DDM) is to increase the training samples. In this paper, an alternative method to reduce the generalization risk was proposed, in which domain knowledges are embedded into the data-driven learning paradigm and it is particularly useful for the scenarios with limited training samples. First, we introduce the formulation of the data-driven model incorporated with knowledge(KDM), and its solutions. Then, the characteristics of the KDM's solutions were analyzed, and the methods to measure KDM's generalization capability in both the global learning space and local learning space were given. Finally, the applicability of KDM in regression analysis, pattern recognition, and dynamic programming problems were discussed. Compared with DDM, the learning procedure in KDM is more efficient and can reduce the generalization risk without increasing the training samples.
引文
[1]Vapnik V N.The nature of statistical learning theory[M].2nd ed.New York:Springer,1999.
[2]Goodfellow I,Bengio Y,Courville A.Deep learning[M].Massachusetts:The MIT Press,2016.
[3]Yeung S,Rinaldo F,Jopling J,et al.A computer vision system for deep learning-based detection of patient mobilization activities in the ICU[J].npj Digital Medicine,2019,2(1):11.
[4]国家电网公司.电网运行风险预警管控工作规范[M].北京:中国电力出版社,2016.State Grid Corporation of China.Specification of operational risk warning and control of power grid[M].Beijing:China Electric Power Press,2016(in Chinese).
[5]国家电网公司.Q/GDW 645-2011配网设备状态评价导则[S].北京:国家电网公司,2011.State Grid Corporation of China.Q/GDW 645-2011Guidelines of condition assessment for electric distribution network equipments[S].Beijing:State Grid Corporation of China,2011(in Chinese).
[6]肖逸思.谷歌兄弟公司Waymo无人驾驶商业化或将推迟[EB/OL].[2019-02-27].http://capital.people.com.cn/n1/2018/1129/c405954-30431545.html.Xiao Yisi.The commercialization of Google’s Waymo driverless car may be delayed[EB/OL].[2019-02-27].http://capital.people.com.cn/n1/2018/1129/c405954-30431545.html(in Chinese).
[7]Besold T R,d'Avila Garcez A,Bader S,et al.Neuralsymbolic learning and reasoning:a survey and interpretation[J].arXiv:1711.03902v1,2017.
[8]中国工程院.全球工程前沿2018[EB/OL].[2019-02-27].http://www.cae.cn/cae/html/main/col1/2018-12/07/20181207171831177603659_1.html.Chinese Academy of Engineering.Engineering fronts2018[EB/OL].[2019-02-27].http://www.cae.cn/cae/html/main/col1/2018-12/07/20181207171831177603659_1.html(in Chinese).
[9]d'Avila Garcez A S,Broda K,Gabbay D M.Neuralsymbolic learning systems:foundations and applications[M].London:Springer,2002.
[10]Ganchev K,Gra?a J,Gillenwater J,et al.Posterior regularization for structured latent variable models[J].The Journal of Machine Learning Research,2010,11:2001-2049.
[11]Hu Zhiting,Ma Xuezhe,Liu Zhengzhong,et al.Harnessing deep neural networks with logic rules[C]//Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics.Berlin,Germany:Association for Computational Linguistics,2016.
[12]Zhang Ao,Li Nan,Pu Jian,et al.τ-FPL:toleranceconstrained learning in linear time[C]//Proceedings of AAAI Conference on Artificial Intelligence.New Orleans,Louisiana,USA:AAAI,2018.
[13]Wang Dayong,Khosla A,Gargeya R,et al.Deep learning for identifying metastatic breast cancer[J].arXiv preprint arXiv:1606.05718,2016.
[14]Christiano P F,Leike J,Brown T B,et al.Deep reinforcement learning from human preferences[C]//Proceedings of Conference on Neural Information Processing Systems.Long Beach,California:NIPS,2017.
[15]Pecka M,Svoboda T.Safe exploration techniques for reinforcement learning-an overview[C]//Proceedings of the 1st International Workshop on Modelling and Simulation for Autonomous Systems.Rome,Italy:Springer,2014:357-375.
[16]Huang J,Wu Fa,Precup D,et al.Learning safe policies with expert guidance[C]//Proceedings of the 32nd International Conference on Neural Information Processing Systems.Montréal,Canada:ACM,2018:9123-9132.
[17]Minervini P,Riedel S.Adversarially regularising neural NLI models to integrate logical background knowledge[C]//Proceedings of the 22nd Conference on Computational Natural Language Learning.Brussels,Belgium:Association for Computational Linguistics,2018.
[18]Wang Hai,Poon H.Deep probabilistic logic:a unifying framework for indirect supervision[J].arXiv preprint arXiv:1808.08485,2018.
[19]Sachan M,Dubey A,Mitchell T,et al.Learning pipelines with limited data and domain knowledge:a study in parsing physics problems[C]//Proceedings of the 32nd International Conference on Neural Information Processing Systems.Montréal,Canada:ACM,2018:140-151.
[20]Tandon N,Mishra B D,Grus J,et al.Reasoning about actions and state changes by injecting commonsense knowledge[J].arXiv preprint arXiv:1808.10012,2018.
[21]Narasimhan M,Lazebnik S,Schwing A G.Out of the box:reasoning with graph convolution nets for factual visual question answering[C]//Proceedings of the 32nd International Conference on Neural Information Processing Systems.Montréal,Canada:ACM,2018:2659-2670.
[22]Liang Xiaodan,Hu Zhiting,Zhang Hao,et al.Symbolic graph reasoning meets convolutions[C]//Proceedings of the 32nd International Conference on Neural Information Processing Systems.Montréal,Canada:ACM,2018:1858-1868.
[23]尚宇炜,马钊,彭晨阳,等.内嵌专业知识和经验的机器学习方法探索(一):引导学习的提出与理论基础[J].中国电机工程学报,2017,37(19):5560-5571.Shang Yuwei,Ma Zhao,Peng Chenyang,et al.Study of a novel machine learning method embedding expertise Part I:proposals and fundamentals of guiding learning[J].Proceedings of the CSEE,2017,37(19):5560-5571(in Chinese).
[24]尚宇炜,马钊,彭晨阳,等.内嵌专业知识和经验的机器学习方法探索(二):引导学习的应用与实践[J].中国电机工程学报,2017,37(20):5852-5861.Shang Yuwei,Ma Zhao,Peng Chenyang,et al.Study of a novel machine learning method embedding expertise(part II):applications and practices of guiding learning[J].Proceedings of the CSEE,2017,37(20):5852-5861(in Chinese).
[25]Watkins C J C H,Dayan P.Technical note:Q-learning[J].Machine Learning,1992,8(3-4):279-292.
[26]Sutton R S,Barto A G.Reinforcement learning:an introduction[M].2nd ed.Cambridge,MA:The MITPress,2018.
[27]Tenenbaum J B,De Silva V,Langford J C.A global geometric framework for nonlinear dimensionality reduction[J].Science,2000,290(5500):2319-2323.
[28]Lei Na,Su Kehua,Cui Li,et al.A geometric view of optimal transportation and generative model[J].arXiv:1710.05488,2017.
[29]Pylvanainen J K,Nousiainen K,Verho P.Studies to utilize loading guides and ANN for oil-immersed distribution transformer condition monitoring[J].IEEETransactions on Power Delivery,2007,22(1):201-207.
[30]Seier A,Hines P D H,Frolik J.Data-driven thermal modeling of residential service transformers[J].IEEETransactions on Smart Grid,2015,6(2):1019-1025.
[31]IEEE.C57.91-2011 Guide for loading mineral-oilimmersed transformers and step-voltage regulators[S].IEEE,2012.
[32]Tan Chuanqi,Sun Fuchun,Kong Tao.A survey on deep transfer learning[C]//Proceedings of the 27th International Conference on Artificial Neural Networks.Rhodes,Greece:Springer,2018.
[33]Liu Chunyang,Wang Xiuli,Wu Xiong,et al.Economic scheduling model of microgrid considering the lifetime of batteries[J].IET Generation,Transmission&Distribution,2017,11(3):759-767.
[34]Liu Weirong,Zhuang Peng,Liang Hao,et al.Distributed economic dispatch in microgrids based on cooperative reinforcement learning[J].IEEE Transactions on Neural Networks and Learning Systems,2018,29(6):2192-2203.
[35]余涛,周斌,甄卫国.强化学习理论在电力系统中的应用及展望[J].电力系统保护与控制,2009,37(14):122-128.Yu Tao,Zhou Bin,Zhen Weiguo.Application and development of reinforcement learning theory in power systems[J].Power System Protection and Control,2009,37(14):122-128(in Chinese).
[2]Goodfellow I,Bengio Y,Courville A.Deep learning[M].Massachusetts:The MIT Press,2016.
[3]Yeung S,Rinaldo F,Jopling J,et al.A computer vision system for deep learning-based detection of patient mobilization activities in the ICU[J].npj Digital Medicine,2019,2(1):11.
[4]国家电网公司.电网运行风险预警管控工作规范[M].北京:中国电力出版社,2016.State Grid Corporation of China.Specification of operational risk warning and control of power grid[M].Beijing:China Electric Power Press,2016(in Chinese).
[5]国家电网公司.Q/GDW 645-2011配网设备状态评价导则[S].北京:国家电网公司,2011.State Grid Corporation of China.Q/GDW 645-2011Guidelines of condition assessment for electric distribution network equipments[S].Beijing:State Grid Corporation of China,2011(in Chinese).
[6]肖逸思.谷歌兄弟公司Waymo无人驾驶商业化或将推迟[EB/OL].[2019-02-27].http://capital.people.com.cn/n1/2018/1129/c405954-30431545.html.Xiao Yisi.The commercialization of Google’s Waymo driverless car may be delayed[EB/OL].[2019-02-27].http://capital.people.com.cn/n1/2018/1129/c405954-30431545.html(in Chinese).
[7]Besold T R,d'Avila Garcez A,Bader S,et al.Neuralsymbolic learning and reasoning:a survey and interpretation[J].arXiv:1711.03902v1,2017.
[8]中国工程院.全球工程前沿2018[EB/OL].[2019-02-27].http://www.cae.cn/cae/html/main/col1/2018-12/07/20181207171831177603659_1.html.Chinese Academy of Engineering.Engineering fronts2018[EB/OL].[2019-02-27].http://www.cae.cn/cae/html/main/col1/2018-12/07/20181207171831177603659_1.html(in Chinese).
[9]d'Avila Garcez A S,Broda K,Gabbay D M.Neuralsymbolic learning systems:foundations and applications[M].London:Springer,2002.
[10]Ganchev K,Gra?a J,Gillenwater J,et al.Posterior regularization for structured latent variable models[J].The Journal of Machine Learning Research,2010,11:2001-2049.
[11]Hu Zhiting,Ma Xuezhe,Liu Zhengzhong,et al.Harnessing deep neural networks with logic rules[C]//Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics.Berlin,Germany:Association for Computational Linguistics,2016.
[12]Zhang Ao,Li Nan,Pu Jian,et al.τ-FPL:toleranceconstrained learning in linear time[C]//Proceedings of AAAI Conference on Artificial Intelligence.New Orleans,Louisiana,USA:AAAI,2018.
[13]Wang Dayong,Khosla A,Gargeya R,et al.Deep learning for identifying metastatic breast cancer[J].arXiv preprint arXiv:1606.05718,2016.
[14]Christiano P F,Leike J,Brown T B,et al.Deep reinforcement learning from human preferences[C]//Proceedings of Conference on Neural Information Processing Systems.Long Beach,California:NIPS,2017.
[15]Pecka M,Svoboda T.Safe exploration techniques for reinforcement learning-an overview[C]//Proceedings of the 1st International Workshop on Modelling and Simulation for Autonomous Systems.Rome,Italy:Springer,2014:357-375.
[16]Huang J,Wu Fa,Precup D,et al.Learning safe policies with expert guidance[C]//Proceedings of the 32nd International Conference on Neural Information Processing Systems.Montréal,Canada:ACM,2018:9123-9132.
[17]Minervini P,Riedel S.Adversarially regularising neural NLI models to integrate logical background knowledge[C]//Proceedings of the 22nd Conference on Computational Natural Language Learning.Brussels,Belgium:Association for Computational Linguistics,2018.
[18]Wang Hai,Poon H.Deep probabilistic logic:a unifying framework for indirect supervision[J].arXiv preprint arXiv:1808.08485,2018.
[19]Sachan M,Dubey A,Mitchell T,et al.Learning pipelines with limited data and domain knowledge:a study in parsing physics problems[C]//Proceedings of the 32nd International Conference on Neural Information Processing Systems.Montréal,Canada:ACM,2018:140-151.
[20]Tandon N,Mishra B D,Grus J,et al.Reasoning about actions and state changes by injecting commonsense knowledge[J].arXiv preprint arXiv:1808.10012,2018.
[21]Narasimhan M,Lazebnik S,Schwing A G.Out of the box:reasoning with graph convolution nets for factual visual question answering[C]//Proceedings of the 32nd International Conference on Neural Information Processing Systems.Montréal,Canada:ACM,2018:2659-2670.
[22]Liang Xiaodan,Hu Zhiting,Zhang Hao,et al.Symbolic graph reasoning meets convolutions[C]//Proceedings of the 32nd International Conference on Neural Information Processing Systems.Montréal,Canada:ACM,2018:1858-1868.
[23]尚宇炜,马钊,彭晨阳,等.内嵌专业知识和经验的机器学习方法探索(一):引导学习的提出与理论基础[J].中国电机工程学报,2017,37(19):5560-5571.Shang Yuwei,Ma Zhao,Peng Chenyang,et al.Study of a novel machine learning method embedding expertise Part I:proposals and fundamentals of guiding learning[J].Proceedings of the CSEE,2017,37(19):5560-5571(in Chinese).
[24]尚宇炜,马钊,彭晨阳,等.内嵌专业知识和经验的机器学习方法探索(二):引导学习的应用与实践[J].中国电机工程学报,2017,37(20):5852-5861.Shang Yuwei,Ma Zhao,Peng Chenyang,et al.Study of a novel machine learning method embedding expertise(part II):applications and practices of guiding learning[J].Proceedings of the CSEE,2017,37(20):5852-5861(in Chinese).
[25]Watkins C J C H,Dayan P.Technical note:Q-learning[J].Machine Learning,1992,8(3-4):279-292.
[26]Sutton R S,Barto A G.Reinforcement learning:an introduction[M].2nd ed.Cambridge,MA:The MITPress,2018.
[27]Tenenbaum J B,De Silva V,Langford J C.A global geometric framework for nonlinear dimensionality reduction[J].Science,2000,290(5500):2319-2323.
[28]Lei Na,Su Kehua,Cui Li,et al.A geometric view of optimal transportation and generative model[J].arXiv:1710.05488,2017.
[29]Pylvanainen J K,Nousiainen K,Verho P.Studies to utilize loading guides and ANN for oil-immersed distribution transformer condition monitoring[J].IEEETransactions on Power Delivery,2007,22(1):201-207.
[30]Seier A,Hines P D H,Frolik J.Data-driven thermal modeling of residential service transformers[J].IEEETransactions on Smart Grid,2015,6(2):1019-1025.
[31]IEEE.C57.91-2011 Guide for loading mineral-oilimmersed transformers and step-voltage regulators[S].IEEE,2012.
[32]Tan Chuanqi,Sun Fuchun,Kong Tao.A survey on deep transfer learning[C]//Proceedings of the 27th International Conference on Artificial Neural Networks.Rhodes,Greece:Springer,2018.
[33]Liu Chunyang,Wang Xiuli,Wu Xiong,et al.Economic scheduling model of microgrid considering the lifetime of batteries[J].IET Generation,Transmission&Distribution,2017,11(3):759-767.
[34]Liu Weirong,Zhuang Peng,Liang Hao,et al.Distributed economic dispatch in microgrids based on cooperative reinforcement learning[J].IEEE Transactions on Neural Networks and Learning Systems,2018,29(6):2192-2203.
[35]余涛,周斌,甄卫国.强化学习理论在电力系统中的应用及展望[J].电力系统保护与控制,2009,37(14):122-128.Yu Tao,Zhou Bin,Zhen Weiguo.Application and development of reinforcement learning theory in power systems[J].Power System Protection and Control,2009,37(14):122-128(in Chinese).
© 2004-2018 中国地质图书馆版权所有 京ICP备05064691号 京公网安备11010802017129号 地址：北京市海淀区学院路29号 邮编：100083 电话：办公室：(+86 10)66554848；文献借阅、咨询服务、科技查新：66554700 |