Rademacher复杂度在统计学习理论中的研究:综述

发布时间：2018-03-30 07:54

本文选题：机器学习　切入点：统计学习理论　出处：《自动化学报》2017年01期

【摘要】：假设空间复杂性是统计学习理论中用于分析学习模型泛化能力的关键因素.与数据无关的复杂度不同,Rademacher复杂度是与数据分布相关的,因而通常能得到比传统复杂度更紧致的泛化界表达.近年来,Rademacher复杂度在统计学习理论泛化能力分析的应用发展中起到了重要的作用.鉴于其重要性,本文梳理了各种形式的Rademacher复杂度及其与传统复杂度之间的关联性,并探讨了基于Rademacher复杂度进行学习模型泛化能力分析的基本技巧.考虑样本数据的独立同分布和非独立同分布两种产生环境,总结并分析了Rademacher复杂度在泛化能力分析方面的研究现状.展望了当前Rademacher复杂度在非监督框架与非序列环境等方面研究的不足,及其进一步应用与发展.
[Abstract]:Assuming the space complexity is the key factor analysis of learning and the generalization ability of the model used in the statistical learning theory. Different complexity and data that is independent of the complexity of Rademacher is associated with the distribution of the data, and usually can be more compact than traditional complex expression of generalization. In recent years, the complexity of Rademacher application development in statistical learning theory the generalization ability plays an important role. In view of its importance, this paper explores the relevance of various forms of Rademacher complexity and traditional complexity, and discusses the Rademacher complexity of basic skills of learning model based on generalization ability. Considering the sample data are independent and identically distributed and non IID two kinds of production environment, summary and analysis of the generalization ability of analysis on the status quo of the Rademacher complex. The prospect of the current Rademacher in complexity The lack of research on unsupervised framework and non sequence environment, and its further application and development.

【作者单位】：上海电子信息职业技术学院计算机应用系;复旦大学计算机科学技术学院;上海市智能信息处理重点实验室;

【相似文献】