[Seminar] Generative models as queryable world models

Abstract: Over the last decade, the de facto standard for training high performing representation learning models has heavily relied on large scale static datasets crawled from the Internet. However, recent advances in visual content creation are challenging this status quo by pushing researchers to leverage high performing image generative models as world models that work in tandem with representation learning models by providing them with data. In this talk, I will address two research questions: (1) Are state-of-the-art image generative models optimized to work as world models?; and (2) What is the most effective way to guide the generative model to produce samples that are useful for the downstream task?

Bio:Adriana Romero-Soriano is currently a research scientist at Meta AI (FAIR), an adjunct professor at McGill University, a core industry member of Mila, and a Canada CIFAR Chair. The playground of her research has been defined by problems which require inferring full observations from limited sensory data, building models of the world with the goal to improve impactful downstream applications responsibly. Her most recent research focuses on improving the quality, consistency, and representation diversity of visual content creation systems. Adriana received her Ph.D. from University of Barcelona, where she worked with Dr. Carlo Gatta, and spent two years as post-doctoral researcher at Mila working with Prof. Yoshua Bengio.