基于支持向量机与概率输出网络的深度学习
发布时间：2019-04-10 19:06
【摘要】：深度学习是能够自动提取特征并且实现对无标签样本学习的深层算法。尽管训练好的深层网络能够提供良好的性能,但是学习算法的超参数需要精确的配置以及人工确定。支持向量机本质上是浅层结构,无法自动提取表征数据的抽象特征。因此在保持支持向量机自身优势的同时,研究具有自动提取数据内部结构特征的能力具有重要的理论意义和实践意义。针对分类问题,论文利用深度学习的深层结构、支持向量机的结构风险最小化以及概率输出网络中的条件概率估计等特点,建立了多层支持向量机结构。其中核参数的选择域呈网格状,通过求取正负两种类别对应β分布的累积概率分布和经验累积概率分布的K-S统计求取一致性的P值乘积最大对应的核参数作为支持向量机模型的核参数。对应的输出为模型提取的特征,作为下一层的输入,直至模型达到结束条件为止。最后通过三个常用的分类数据集对所提模型进行了实验验证和分析。针对回归问题,论文利用深度学习的深层结构、支持向量机的结构风险最小化以及概率输出网络中的条件概率估计等特点,建立了多层支持向量机结构。其中核参数的选择域呈网格状,通过求取输出对应β分布的累积概率分布和经验累积概率分布的K-S统计求取一致性的P值最大对应的核参数作为支持向量机模型的核参数。对应的输出为模型提取的特征,作为下一层的输入,直至模型达到结束条件为止。最后通过三个常用的回归数据集对所提模型进行了实验验证和分析。
[Abstract]:Deep learning is a deep algorithm which can automatically extract features and realize unlabeled sample learning. Although the trained deep network can provide good performance, the hyperparameters of the learning algorithm need precise configuration and manual determination. Support vector machine (SVM) is essentially a shallow structure and can not automatically extract the abstract features of the representation data. Therefore, while maintaining the advantages of SVM itself, it is of great theoretical and practical significance to study the ability of automatically extracting the internal structural features of data. Based on the deep structure of deep learning, structural risk minimization of support vector machines and conditional probability estimation in probability output networks, a multi-layer support vector machine structure is proposed in this paper. Where the selection domain of the nuclear parameters is grid-shaped, By calculating the cumulative probability distribution of the positive and negative classes corresponding to the 尾 distribution and the K S statistics of the empirical cumulative probability distribution, the kernel parameters corresponding to the maximum product of the consistent P value are obtained as the kernel parameters of the support vector machine model. The corresponding output is the feature extracted from the model, which is used as the input of the next layer until the end condition of the model is reached. Finally, three common classification data sets are used to verify and analyze the proposed model. Based on the deep structure of deep learning, structural risk minimization of support vector machines and conditional probability estimation in probability output networks, a multi-layer support vector machine structure is proposed in this paper. The kernel parameter selection domain is grid-shaped, and the kernel parameters corresponding to the consistent P value are obtained by calculating the cumulative probability distribution of the output 尾 distribution and the empirical cumulative probability distribution as the kernel parameters of the support vector machine model. The corresponding output is the feature extracted from the model, which is used as the input of the next layer until the end condition of the model is reached. Finally, three commonly used regression data sets are used to verify and analyze the proposed model.
【学位授予单位】：西安理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP181
本文编号：2456044
[Abstract]:Deep learning is a deep algorithm which can automatically extract features and realize unlabeled sample learning. Although the trained deep network can provide good performance, the hyperparameters of the learning algorithm need precise configuration and manual determination. Support vector machine (SVM) is essentially a shallow structure and can not automatically extract the abstract features of the representation data. Therefore, while maintaining the advantages of SVM itself, it is of great theoretical and practical significance to study the ability of automatically extracting the internal structural features of data. Based on the deep structure of deep learning, structural risk minimization of support vector machines and conditional probability estimation in probability output networks, a multi-layer support vector machine structure is proposed in this paper. Where the selection domain of the nuclear parameters is grid-shaped, By calculating the cumulative probability distribution of the positive and negative classes corresponding to the 尾 distribution and the K S statistics of the empirical cumulative probability distribution, the kernel parameters corresponding to the maximum product of the consistent P value are obtained as the kernel parameters of the support vector machine model. The corresponding output is the feature extracted from the model, which is used as the input of the next layer until the end condition of the model is reached. Finally, three common classification data sets are used to verify and analyze the proposed model. Based on the deep structure of deep learning, structural risk minimization of support vector machines and conditional probability estimation in probability output networks, a multi-layer support vector machine structure is proposed in this paper. The kernel parameter selection domain is grid-shaped, and the kernel parameters corresponding to the consistent P value are obtained by calculating the cumulative probability distribution of the output 尾 distribution and the empirical cumulative probability distribution as the kernel parameters of the support vector machine model. The corresponding output is the feature extracted from the model, which is used as the input of the next layer until the end condition of the model is reached. Finally, three commonly used regression data sets are used to verify and analyze the proposed model.
【学位授予单位】：西安理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP181
【参考文献】
相关期刊论文 前7条
1 段艳杰;吕宜生;张杰;赵学亮;王飞跃;;深度学习在控制领域的研究现状与展望[J];自动化学报;2016年05期
2 张仕光;胡清华;谢宗霞;米据生;;基于Beta噪声模型支持向量回归及其应用[J];南京大学学报(自然科学版);2013年04期
3 陈进东;潘丰;;基于在线支持向量机和遗传算法的预测控制[J];系统工程与电子技术;2013年06期
4 崔和;龙玉峰;;支持向量机学习算法的研究现状与展望[J];信息与电子工程;2008年05期
5 刘涵;刘丁;;基于支持向量机的参数自整定PID非线性系统控制[J];控制理论与应用;2008年03期
6 郭小荟;马小平;;基于Matlab的支持向量机工具箱[J];计算机应用与软件;2007年12期
7 张学工;关于统计学习理论与支持向量机[J];自动化学报;2000年01期
，本文编号：2456044
本文链接：https://www.wllwen.com/kejilunwen/zidonghuakongzhilunwen/2456044.html
