Towards Improving Single Label and Hierarchical Multi-Label Classification

Teaching a computer to do what humans can do is the ultimate goal of artificial intelligence. In the area of artificial intelligence, classification is the most fundamental and important topic in machine learning, data mining, and pattern recognition. Humans can unconsciously distinguish different objects or patterns, then classify them. However, it is not easy to tell computer program to do the same. To this end, different classification algorithms have been developed and applied to different domains, such as computer vision, natural language processing, speech recognition, genetic engineering, and financial markets. Therefore, improving the performance of classification algorithms is crucially important for all these applications.

The goal of this research is to develop new classification algorithms that improve the performance of single-label classification for visual recognition and hierarchical multi-label classification, and apply the developed algorithms in the UR2D system. The specific objectives in single-label classification are to: (i) overcome the challenge of increasing deep-neural-network depth; (ii) overcome the challenge of occlusion in 2D-2D face recognition; (iii) overcome the challenge of using soft-facial attribute features in 2D-2D face recognition. The specific objective in hierarchical multi-label classification is to take advantage of the complex label correlations in the class hierarchy. The proposed algorithms were evaluated on miscellaneous datasets and achieved significant improvements when compared with previous baseline methods.