Constructing training distribution by minimizing variance of risk criterion for visual category learning

This work addresses the problem of constructing an effective training set at minimal labeling cost by selecting some images to build a subset from the whole database. This problem occurs in situations that the number of categories is large or the cost of obtaining labeled images is extremely high, because the images selected by uniform sampling do not reflect the desired training distribution and need additional labeling cost in order to obtain enough labeled images. We study the active training process in which the images are actively sampled process from a pool of unlabeled images, and then their labels are queried. We construct a training distribution by minimizing variance of structural risk of classification model. The experimental results show that our approach can derive more accurate model than the existing method with the same labeling cost, and our approach is proved to be more effective than common uniform sampling in which the images are drawn equally from the whole database.