February 14th

14:00, Shannon amphitheatre (building 660) (see location):

Victor Berger (Thales Services, ThereSIS)


Title: VAE/GAN as a generative model


Abstract:


We investigate the problem of data generation, i.e., the unsupervised training of a model to generate samples from a distribution generalizing a dataset. We use from [1] an approach combining the Variational Autoencoder (VAE) [2] model with the well-known Generative Adversarial Network (GAN) [3]. As observed in recent literature, training a GAN model is tedious and subject to instability in the optimization process. We reproduce results from [1] and explore different architectures and techniques for taming these instabilities.

In this presentation, we first introduce the VAE and GAN models. Then we detail the approach from [1], and provide experimental results in favor of the following conclusions: combining VAE and GAN stabilizes the training and induces smoothness in the latent space of the generative network, while keeping the sharpness of the generated images.

[1] Larsen, A. B. L., Sønderby, S. K., Larochelle, H., & Winther, O. (2015). Autoencoding beyond pixels using a learned similarity metric. arXiv preprint arXiv:1512.09300.
[2] Kingma, D. P., & Welling, M. (2013). Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114.
[3] Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., & Bengio, Y. (2014). Generative adversarial nets. In Advances in neural information processing systems (pp. 2672-2680).




Contact: guillaume.charpiat at inria.fr