Rotational weight update in full-connection layers exceeds dropout in image recognition tasks

Various methods have been proposed to improve image recognition performance with deep learning. In this study, we propose a method named “Rotational-update” that updates the weights of neurons in rotation. The method divides the neurons into $\sqrt{N}$ groups of equal size, where N is the number of neurons in a full-connection layer. The method updates the neuron weights of only one group in turn for each mini-batch. One important feature is that this method can be combined with other methods such as bach-normalization and dropout.We verified the effectiveness of the proposed method through the image recognition task of a dataset CIFARIO with a neural network model VGG-16. The results showed that combining the proposed method and bach-normalization was the most accurate and better than the combination of dropout and bach-normalization. We conclude that using the proposed method instead of dropout improves the performance of image recognition tasks.The proposed method is provided as a library on pip (pip install rotational-update).