An attempt of finding an appropriate number of convolutional layers in CNNs based on benchmarks of heterogeneous datasets

Abstract

An attempt of finding an appropriate number of convolutional layers in convolutional neural networks is made. The benchmark datasets are CIFAR-10, NORB, and EEACL26, whose diversity and heterogeneousness must serve for a general applicability of a rule presumed to give that number. The rule is drawn from the best performances of convolutional neural networks built with 2 to 12 convolutional layers. It is not an exact best number of the convolutional layers, but a short process of trying a few versions of those numbers. For small images (like those ones in CIFAR-10), that initial number is 4. For datasets having a few tens of image categories and more, initially setting 5 to 8 convolutional layers is recommended depending on the dataset complexity. The fuzziness in the rule is not removable because of the required diversity and heterogeneousness.