Mahdi Azarafrooz

Feb 15, 2018 (modified: Feb 15, 2018) ICLR 2018 Conference Blind Submission readers: everyone Show Bibtex
  • Abstract: Any autoencoder network can be turned into a generative model by imposing an arbitrary prior distribution on its hidden code vector. Variational Autoencoder uses a KL divergence penalty to impose the prior, whereas Adversarial Autoencoder uses generative adversarial networks. A straightforward modification of Adversarial Autoencoder can be achieved by replacing the adversarial network with maximum mean discrepancy (MMD) network. This replacement leads to a new set of probabilistic autoencoder which is also discussed in our paper. However, an essential challenge remains in both of these probabilistic autoencoders, namely that the only source of randomness at the output of encoder, is the training data itself. Lack of enough stochasticity can make the optimization problem non-trivial. As a result, they can lead to degenerate solutions where the generator collapses into sampling only a few modes. Our proposal is to replace the adversary of the adversarial autoencoder by a space of {\it stochastic} functions. This replacement introduces a a new source of randomness which can be considered as a continuous control for encouraging {\it explorations}. This prevents the adversary from fitting too closely to the generator and therefore leads to more diverse set of generated samples. Consequently, the decoder serves as a better generative network which unlike MMD nets scales linearly with the amount of data. We provide mathematical and empirical evidence on how this replacement outperforms the pre-existing architectures.
  • Keywords: Generative adversarial Networks, Deep Generative models, Kernel Methods