Export Reviews, Discussions, Author Feedback and Meta-Reviews

Summary: This paper proposed a novel approach for efficient metric learning. The objective of classic metric learning is to solve the positive semi-definite matrix M such that the distance between examples from the same class get closer while the distance between examples from different classes get further in the distance define by M. Unlike these methods, this paper proposed a idea of virtual point-based regression formulation. The points from the same class are pulled toward the virtual points, thus their distance to each other become closer and their distance to points in other classes become further. The paper describes a closed form solution to compute the regression matrix and two different ways to discover the virtual points. The method is evaluated in 13 different metric learning datasets and compared with several standard baselines.