Attribute-Based Synthetic Network (ABS-Net): Learning more from pseudo feature representations. (August 2018)

• Record Type: Journal Article
• Authors: Lu, Jiang; Li, Jin; Yan, Ziang; Mei, Fenghua; Zhang, Changshui
• Is Part Of: Pattern recognition. Volume 80(2018:Aug.)
• Page Start: 129
• Page End: 142
• Publication Date: 2018-08
• Subjects: Pseudo feature representation -- Zero-shot learning -- Supervised learning -- Data augmentation -- Attribute learning; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2018.03.006
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: ABS-Net is proposed to deal with the lack of annotated samples. ABS-Net fills the gaps between seen and unseen concepts. ABS-Net achieves competitive results on some zero-shot learning benchmark datasets. ABS-Net realizes feature-level data augmentation for supervised learning. Experiments on C-MNIST demonstrate the effectiveness of the proposed model. Abstract: In large-scale visual recognition tasks, researchers are usually faced with some challenging problems, such as the extreme imbalance in the number of training data between classes or the lack of annotated data for some classes. In this paper, we propose a novel neural network architecture that automatically synthesizes pseudo feature representations for the classes in lack of annotated images. With the supply of semantic attributes for classes, the proposed Attribute-Based Synthetic Network (ABS-Net) can be applied to zero-shot learning (ZSL) scenario and conventional supervised learning (CSL) scenario as well. For ZSL tasks, the pseudo feature representations can be viewed as annotated feature-level instances for novel concepts, which facilitates the construction of unseen class predictor. For CSL tasks, the pseudo feature representations can be viewed as products of data augmentation on training set, which enriches the interpretation capacity of CSL systems. We demonstrate the effectiveness of the proposed ABS-Net in ZSL and CSL settings on a synthetic colored MNIST dataset (C-MNIST). For several popular ZSL benchmark datasets, our architecture also shows competitive results on zero-shot recognition task, especially leading to tremendous improvement to state-of-the-art mAP on zero-shot retrieval task.