Multigradient-based Convolutional Neural Network

Abstract

The convolutional neural network (CNN) is a promising algorithm for artificial intelligence. Although it was developed for image classification, much research is currently in progress in various fields, such as object detection and image processing. The basic principle of the CNN, especially for classification, is to adopt a loss function and minimize it in an iterative way. In this paper, a multigradient-based training algorithm is proposed for image classification. The proposed algorithm defines an object function based on multigradients and trains the CNN by maximizing the corresponding objective function. When applied to open access databases, the proposed algorithm performed better than conventional back-propagation based CNN methods.