Comparison of the error rates on the CIFAR-10 dataset (small-data scenario). We run the experiment ten times and report the classification errors in the format of “best (mean $±$ std).” The numbers of learnable weight parameters of the best architecture are also reported.
VGG (Simonyan and Zisserman, 2015) $23.05(24.07±0.43)$ 15.2M
ResNet (He et al., 2016) $24.01(24.83±0.51)$ 1.70M
CGP-CNN (ConvSet) $19.78(22.14±1.80)$ 1.94M
CGP-CNN (ResSet) $19.33(20.52±1.36)$ 0.92M
VGG (Simonyan and Zisserman, 2015) $23.05(24.07±0.43)$ 15.2M
ResNet (He et al., 2016) $24.01(24.83±0.51)$ 1.70M
CGP-CNN (ConvSet) $19.78(22.14±1.80)$ 1.94M
CGP-CNN (ResSet) $19.33(20.52±1.36)$ 0.92M