Generative Adversarial Reservoirs for natural video prediction

In this report, we will give a brief overview of selected deep learning technologies in the interest of developing both understanding and motivation for the use of reservoir computing and generative models. Furthermore, we will show that these concepts can be applied to the problem of natural video prediction. Influenced by previous work, we develop a novel architecture called Generative Adversarial Reservoirs (GAR). We use GARs to predict frames of videos from the UCF-101 dataset and show that although some of the quantitative evaluations for our results are below state-of-the-art, utilizing reservoirs allows our model training to converge significantly faster while still achieving qualitatively good results.