Learning to Transfer Privileged Ranking Attribute for Object Classification
Shengzheng Wang · Journal of Information and Computational Science · 2015
Learning Using Privileged Information (LUPI) provides an efiective framework to solve the learning problem under situation of the asymmetric distribution of information between training and test time. It has been successfully applied in the category recognition, e.g., protein classiflcation, hand-writing recognition, animal categorization, etc. However, in the existing methods, various semantic attributes, with the help of experts, were only simply translated into the feature vectors and considered as the privileged data, which restricts the LUPI to the simple applications since it is di‐cult to guarantee that the privileged data is similarly informative about the problem at hand as the original data. Therefore, this paper presents a novel approach based on an attribute-ranking learning algorithm to construct the example-oriented privileged data. The main idea is to provide an efiective means to transfer the midlevel semantic attributes to the original training data. Namely, we flrst obtain a real-valued rank per attribute for each example indicating the relative strength of the attribute presence in all examples, and then the resulting attribute ranking results are used to generate the privileged data. The experimental results show that the proposed approach provides a promising means to apply the privileged ranking attributes, and further demonstrate signiflcant improvements in classiflcation accuracy on three typical databases: PubFig, OSR and AwA.