This study focuses on supplementing data sets with data of absent classes by using other, similar data sets in which these classes are represented. The data is generated using Generative Adversarial Nets (GANs) trained on the CelebA and MNIST datasets. In particular we use and compare Coupled GANs (CoGANs), Auxiliary Classifier GANs (AC-GANs) and novel a combination of the two (CoAC-GANs) to generate image data of domain-class combinations that were removed from the training data. We also train classifiers on the generated data. The results show that AC-GANs and CoAC-GANs can be used successfully to generate labeled data from domain-class combinations that are absent from the training data. Furthermore, they suggest that the preference for one of the two types of generative models depends on training set characteristics. Classifiers trained on the generated data can accurately classify unseen data from the missing domain-class combinations.
Luuk Boulogne, Klaas Dijkstra and Marco Wiering
Published at the IEEE Symposium Series on Computational Intelligence, SCCI 2018