Exploring and exploiting hierarchical structures for large-scale classification

Classification and recognition tasks confronted by intelligent systems are becoming complicated as the sizes of samples, dimensionality and labels dramatically increase in the past few years. Learning machines have to deal with modeling tasks with millions of samples and ten of thousand of labels in high-dimensional feature spaces. These tasks pose great challenges for traditional learning paradigms. Inspired by the progress in cognitive and neural science, we propose an end-to-end deep hierarchical classification framework and integrate the processes of representation learning, hierarchical structure construction, and hierarchical classification modeling in this work. This integrative framework brings several advantages: (1) the learned features are good for hierarchical structure construction and classification; (2) the learned hierarchical structure is beneficial for the subsequent hierarchical classification; (3) the learned hierarchical classification model supervises and guides the representation learning and hierarchical structure construction. This new hierarchical paradigm can not only well deal with large-scale classification tasks but also provide new inspirations to other research fields of artificial intelligence.