Synchronous Dropout for Convolutional Neural Network
Ikkei Sakurai, Chihiro Ikuta · 2021
In this paper, we propose a synchronous dropout that improves convolutional neural network (SD-CNN). The dropout is famous normalization technique for the artificial neural network. In the standard dropout, the neurons are randomly chosen with a fixed probability. In our method, we decompose the neurons in hidden layer to some groups. The neurons in hidden layer are randomly chosen by neuron groups, and all neurons in chosen group, are skipped a training phase with iteration. The group dropout increases training normalization than the standard dropout because the number of chosen neurons is constant. We compare that the group dropout with the standard dropout by using Fashion-MNIST.