期刊名称：IAENG International Journal of Computer Science
印刷版ISSN：1819-656X
电子版ISSN：1819-9224
出版年度：2019
卷号：46
期号：2
页码：365-376
出版社：IAENG - International Association of Engineers
摘要：Convolutional Neural Network (CNN) classifier isa very popular classifier used to solve many problems, includingimage classification and object recognition. The CNN classifierusually improved by designing a deeper and bigger classifierwhich needs more memory and computational power to runthe classifier. In this paper, we analyze and optimize the use ofsmall and shallow CNN classifier on CIFAR dataset. KarpathyConvNetJS CIFAR10 model was used as a base network of ourclassifier and extended by adding max-min pooling method.The max-min pooling is used to explore the negative andpositive response of the convolution process which in theorywill be trained the classifier more effectively. We choose severaldifferent configurations to analyze the effectiveness of theclassifier by combining the training algorithm, batch normalizationconfiguration, weights initialization methods, dropoutregularization configuration, and heavy data augmentation. Toensure that the classifier we designed is still small and shallowCNN classifier, we limit the maximum number of layers inour CNN classifier to 15 layers. Experiments on CIFAR10 andCIFAR100 dataset shows that by compacting the kernel on eachlayer, the classifier can achieve good accuracy and comparablewith another state-of-the-art classifier with a relatively samenumber of layers with an error rate of 6.99% on the CIFAR10dataset and 29.41% on the CIFAR100 dataset.
