Deep Generative Models

One of the biggest challenges for research in artificial intelligence is unsupervised learning. Current industrial success with deep learning relies heavily on supervised learning, where humans are needed to categorize data and define high-level abstractions which we want the computer to know about. However, humans are able to discover many aspects of the world without a teacher telling them anything about it, and this ability to autonomously learn to make sense of the world is something that needs to be further developed for computers. The deep learning approach to unsupervised learning centers on the question of learning representations, and different algorithms define an objective function which leads the learner to capture essential aspects of the data distribution along with a new space in which to represent data. Deep generative models can demonstrate their understanding of the data by generating novel examples which nonetheless look like those which were used to train the model. Many of these models are related to the old idea of auto-encoder, with an encoder function mapping data to representation and a decoder (or generator) mapping the abstract representation to the raw data space. The talk will focus in particular on the family of generative adversarial networks (GANs), which question the established approaches based on maximum likelihood and probability function estimation and bring us into the realm of game theory and novel ways of comparing different distributions against each other, as well as with very impressive generation of images.