Activation Maximization Generative Adversarial Nets
share
Class label information has been empirically proven to be very useful in improving the sample quality of generative adversarial nets (GANs). In this paper, we mathematically study current variants of GANs that make use of class label information to reveal how class labels and associated losses influence GAN's training. Based on the analysis, we propose Activation Maximization Generative Adversarial Networks (AM-GAN) as an alternative solution. We conduct a set of controlled experiments to validate our analysis and study the effectiveness of our solution, where AM-GAN achieves the state-of-the-art Inception Score (8.91) on CIFAR-10. Through the experiments, we realize the common used metric for generative models needs further investigation and refinement. Thus we also delve into the widely-used evaluation metrics and accordingly propose a new metric as compensation to make the entire metrics complete and impartial. The proposed model also outperforms the baseline methods in the new metric.
READ FULL TEXT