AI Seminar sponsored by Apple

Guoqing Zheng -- Generative Adversarial Permutation Learning

Abstract: Permutation learning refers to the task of recovering the mapping that permutes an object to another one. The (original object, permuted object) pair encodes information that is critical in identifying the underlying permutation and has received increasingly attention in the research community. In this talk, we study the problem of unpaired permutation learning, i.e., only samples of original objects and that of permuted objects are observed while no paired link between the two is given. We propose to tackle the unpaired permutation learning under the adversarial training framework; specifically, a permutation generative network is trained to generate approximated permutations conditioned on the permuted object, by which the permuted object can be transformed back to the original space, and a discriminative network is trained to distinguish real objects from the original space and the recovered ones. Preliminary empirical experiments on sorting numbers and recovering scrambled images demonstrates the effectiveness of the proposed method.