PhD Seminar: Bridging the Gap between Generative Adversarial Networks (GANs) and Varitional Autoencoders (VAE)
I will present an overview of modern generative models in deep learning. In particular, I will consider GANs and VAE models as the most successful methods to generate very high-dimensional objects such as image, audio, video and so on. While these models are very powerful, they have their own drawbacks. It is known that GANs can produce very realistic samples, but they suffer from the mode collapsing problem when they end up generating only a small subset of real samples. At the same time, VAE is directly optimized to cover all real samples at the cost of covering low-probability regions as well, i.e., it can generate unrealistic samples. I will share our results on combining the best properties of GANs and VAEs. I will describe our proposed generative model, its current performance, and ideas how we can improve it.