Training bidirectional generative adverarial networks with hints

Abstract

The generative adversarial network (GAN) is a deep learning architecture that learns a generative model by training a later discriminator to best differentiate “fake” examples generated by the generator from the “true” examples sampled from the train ing set. The generator of GAN takes a low-dimensional latent space vector as input and learns to generate the corresponding input example. The aim of the generator is to gen erate examples that can not be separated from the true examples by the discriminator. The aim of the discriminator is to maximize the separability of the generated exam ples from the true examples. A recent extension is the bidirectional GAN (BiGAN) where an encoder is also trained in the inverse direction to generate the latent space vector for a given training example. Recently, Wasserstein GAN has been proposed for GAN and our first contribution is to adapt Wasserstein loss to BiGANs. The added encoder of the BiGAN also allows us to define auxiliary reconstruction losses as hints to learn a better generator, and this is our second contribution. Through experiments on five image data sets, namely, MNIST, UT-Zap50K, GTSRB, Cifar10, and CelebA, we show that Wasserstein BiGANs, augmented with hints, learn better generators in terms of image generation quality and diversity, as measured visually by analyzing the generated samples, and numerically by the 1-nearest-neighbor test.