Pablo M. Olmos, Briland Hitaj, Paolo Gasti, Giuseppe Ateniese, Fernando Perez-Cruz

Feb 15, 2018 (modified: Feb 15, 2018) ICLR 2018 Conference Blind Submission readers: everyone Show Bibtex
  • Abstract: GANs have shown how deep neural networks can be used for generative modeling, aiming at achieving the same impact that they brought for discriminative modeling. The first results were impressive, GANs were shown to be able to generate samples in high dimensional structured spaces, like images and text, that were no copies of the training data. But generative and discriminative learning are quite different. Discriminative learning has a clear end, while generative modeling is an intermediate step to understand the data or generate hypothesis. The quality of implicit density estimation is hard to evaluate, because we cannot tell how well a data is represented by the model. How can we certainly say that a generative process is generating natural images with the same distribution as we do? In this paper, we noticed that even though GANs might not be able to generate samples from the underlying distribution (or we cannot tell at least), they are capturing some structure of the data in that high dimensional space. It is therefore needed to address how we can leverage those estimates produced by GANs in the same way we are able to use other generative modeling algorithms.
  • Keywords: Generative Modeling, Generative Adversarial Networks, Density Estimation
0 Replies