HIERARCHICAL QUADRUPLET NET FOR DEEP METRIC LEARNING AND NETWORK REGULARIZATION

Deep neural networks (DNNs) have been used successfully for various artificial intelligence tasks. However, most of these DNN models are only used for classification or regression task. In this paper, Hierarchical Quadruplet Net (HQN) is proposed to map images into Euclidean space, which can be applied for tasks based on distance comparison, such as image retrieval, clustering and classification. HQN uses more prior knowledge and performs better than previous deep metric learning methods, achieving 94.4% mean average precision (MAP) on CIFAR-10 dataset and 76.1% MAP on CIFAR-100 dataset. Moreover, the proposed loss function, Hierarchical Quadruplet Loss (HQL) also improves classification accuracy by 0.1-2.2%.