【学术报告】美国中佛罗里达大学齐国君助理教授于6月16日来访实验室

In this talk, I will present a novel Loss-Sensitive GAN (LS-GAN) that learns a loss function to separate generated samples from their real examples.  An important property of the LS-GAN is it allows the generator to focus on improving poor data points that are far apart from real examples rather than wasting efforts on those samples that have already been well generated, and thus can improve the overall quality of generated samples. The theoretical analysis also shows that the LS-GAN can generate samples following the true data density. In particular, we present a regularity condition on the underlying data density, which allows us to use a class of Lipschitz losses and generators to model the LS-GAN.  It relaxes the assumption that the classic GAN should have infinite modeling capacity to obtain the similar theoretical guarantee. Furthermore, we show the generalization ability of the LS-GAN by bounding the difference between the model performances over the empirical and real distributions, as well as deriving a tractable sample complexity to train the LS-GAN model in terms of its generalization ability. We also derive a non-parametric solution that characterizes the upper and lower bounds of the losses learned by the LS-GAN, both of which are cone-shaped and have non-vanishing gradient almost everywhere. This shows there will be sufficient gradient to update the generator of the LS-GAN even if the loss function is over-trained, relieving the vanishing gradient problem in the classic GAN.    
We also extend the unsupervised LS-GAN to a conditional model generating samples based on given conditions, and show its applications in both supervised and semi-supervised learning problems.    
We conduct experiments to compare different models on both generation and classification tasks, and show the LS-GAN is resilient against vanishing gradient and model collapse even with overtrained loss function or mildly changed network architecture.