Annealing Genetic GAN for Imbalanced Web Data Learning

Class imbalance is one of the most basic and important problems web data. The key to overcoming class increase effective instances minority, that is, data augmentation. Generative Adversarial Networks (GANs), which have recently been successfully applied in field image generation, can be used for augmentation because they learn distribution given ample training generate more However, learning distributions from imbalanced make GANs easily get stuck a local optimum. In this work, we propose new strategy called Annealing Genetic GAN (AGGAN), incorporates simulated annealing genetic algorithm into process GANs. And help avoid optimum trapping problem, occurs when set imbalanced. Unlike existing GANs, use fixed adversarial objective alternately generator, multiple objectives train generators Metropolis criterion decide whether generator should update. More specifically, accepts worse solutions with certain probability, so it our AGGAN escape find better solution. Theory mathematical analysis provide strong theoretical support proposed strategy. experiments on several datasets demonstrate achieves convincing ability solve problem reduces inherent