ImageNet as a Representative Basis for Deriving Generally Effective CNN Architectures

Lukas Tuggener, Thilo Stadelmann, Jürgen Schmidhuber

Keywords: ImageNet, CNN design, dataset representativeness, empirical study

Abstract: We investigate and improve the representativeness of ImageNet as a basis for deriving generally effective convolutional neural network (CNN) architectures that perform well on a diverse set of datasets and application domains. To this end, we conduct an extensive empirical study for which we train 500 CNN architectures, sampled from the broad AnyNetX design space, on ImageNet as well as 8 other image classification datasets. We observe that the performances of the architectures are highly dataset dependent. Some datasets even exhibit a negative error correlation with ImageNet across all architectures. We show how to significantly increase these correlations by utilizing ImageNet subsets restricted to fewer classes. We also identify the cumulative width across layers as well as the total depth of the network as the most sensitive design parameter with respect to changing datasets.

One-sentence Summary: Based on an extensive empirical study we investigate the representativeness of ImageNet as a basis for generally effective CNN architectures and show how to increase said representativeness using class wise downsampling