机器学习泛化能力在小型数据库上的鲁棒性研究  

Robustness of Machine Learning Models for Small Structured Datasets

作　　者：许峻彬 张勇 赵鸿 XU Jun-bin;ZHANG Yong;ZHAO Hong(School of Physical Science and Technology,Xiamen University,Xiamen,Fujian,361005,China)

机构地区：[1]厦门大学物理科学与技术学院,福建厦门361005

出　　处：《新疆师范大学学报（自然科学版）》2021年第2期61-66,共6页Journal of Xinjiang Normal University(Natural Sciences Edition)

基　　金：国家自然科学基金项目(11975189)。

摘　　要：机器学习算法的鲁棒性对于实际应用十分重要。大多相关研究主要针对大型数据库上的深度模型,针对小型数据库的研究还较为缺乏,困难在于小型数据集上尚无有效的方法来实际计算鲁棒性。文章引入了一种适合小数据集的鲁棒性计算方法,并利用Coimbra乳腺癌数据库评估了多种算法的鲁棒性,发现在训练过程中能有效控制输入输出敏感性的算法具有较强的鲁棒性,同时也具有较高的平均泛化能力。The robustness of machine learning algorithm is very important for practical application.Most of the relevant studies focus on the deep models on large databases,while the research on small databases is still lacking.The difficulty is that there is no effective method to calculate robustness on small datasets.This paper introduces a new method to calculate robustness for small datasets and uses Coimbra breast cancer database to evaluate the ro-bustness of various algorithms.It is found that the algorithm which can effectively control input-output sensitivity in training process has strong robustness and high average generalization ability.

关 键 词：机器学习 神经网络 泛化能力 鲁棒性 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]