Training Group Orthogonal Neural Networks with Privileged Information

Yunpeng Chen, Xiaojie Jin, Jiashi Feng, Shuicheng Yan

Abstract: Learning rich and diverse feature representation are always desired for deep convolutional neural networks (CNNs). Besides, when auxiliary annotations are available for specific data, simply ignoring them would be a great waste. In this paper, we incorporate these auxiliary annotations as privileged information and propose a novel CNN model that is able to maximize inherent diversity of a CNN model such that the model can learn better feature representation with a stronger generalization ability. More specifically, we propose a group orthogonal convolutional neural network (GoCNN) to learn features from foreground and background in an orthogonal way by exploiting privileged information for optimization, which automatically emphasizes feature diversity within a single model. Experiments on two benchmark datasets, ImageNet and PASCAL VOC, well demonstrate the effectiveness and high generalization ability of our proposed GoCNN models.

TL;DR: A convolutional neural network for image classification which encourages learning more diverse feature representations by using image segmentations as privileged information.

Conflicts: u.nus.edu, nus.edu.sg, 360.cn

Keywords: Deep learning, Computer vision, Supervised Learning