Grouped Pointwise Convolutions Reduce Parameters in Convolutional Neural Networks

In DCNNs, the number of parameters in pointwise convolutions rapidly grows due to multiplication filters by input channels that come from previous layer. Our proposal makes parameter efficient via grouping into parallel branches or groups, where each branch processes a fraction channels. However, doing so, learning capability DCNN is degraded. To avoid this effect, we suggest interleaving output different at intermediate layers consecutive convolutions. We applied our improvement EfficientNet, DenseNet-BC L100, MobileNet and V3 Large architectures. trained these architectures with CIFAR-10, CIFAR-100, Cropped-PlantDoc The Oxford-IIIT Pet datasets. When training scratch, obtained similar test accuracies original EfficientNet while saving up 90% 63% flops.