The primal goal of generative models is learning a density estimator for a given data like sound, pictures and sentences.

 Thanks to the recently appeared bridge between deep learning and machine learning, many conventional machine learning problems have been solved with astonishing results.

 We have explored generative models in various aspects to get well generalized realistic samples from data.

 Any learning algorithm depends on its underlying assumptions about data, otherwise known as its inductive bias.

 We are developing methods that can learn such assumptions from related datasets, thus creating learners that can (meta-) learn their own learning algorithms.