摘要
分类器模型之间的多样性是分类器集成的一个重要性能指标。目前大多数多样性度量方法都是基于基分类器模型的0/1输出结果(即Oracle输出)进行计算,针对卷积神经网络的概率向量输出结果,仍需要将其转化为Oracle输出方式进行度量,这种方式未能充分利用卷积神经网络输出的概率向量所包含的丰富信息。针对此问题,利用多分类卷积神经网络模型的输出特性,提出了一种基于卷积神经网络的概率向量输出方式的集成多样性度量方法,建立多个不同结构的卷积神经网络基模型并在CIFAR-10和CIFAR-100数据集上进行实验。实验结果表明,与双错度量、不一致性度量和Q统计多样性度量方法相比,所提出的方法能够更好地体现模型之间的多样性,为模型选择集成提供更好的指导。
Diversity among classifier models has been recognized as a significant performance index of a classifier ensemble.Currently,most diversity measuring methods are defined based on the 0/1 outputs(namely Oracle outputs)of the base model.The probability vector outputs of a convolutional neural network(CNN)still need to be converted into Oracle outputs for measurement,which fails to fully use the rich information contained in the CNN probability vector outputs.To solve this problem,a new diversity measuring method for probabilistic vector outputs based on CNNs is proposed.Several base models of CNN models with various structures are established and tested on the CIFAR-10 and CIFAR-100 datasets.Compared with double-fault measure,disagreement measure,and Q-Statistic,the proposed method can better reflect the differences between the models and provide better guidance for a selective ensemble of CNN models.
作者
汤礼颖
贺利乐
何林
屈东东
TANG Liying;HE Lile;HE Lin;QU Dongdong(School of Mechanical and Electrical Engineering,Xi’an University of Architecture and Technology,Xi’an 710055,China;School of Science,Xi’an University of Architecture and Technology,Xi’an 710055,China)
出处
《智能系统学报》
CSCD
北大核心
2021年第6期1030-1038,共9页
CAAI Transactions on Intelligent Systems
基金
国家自然科学基金项目(61903291)。