I work in the broad area of probabilistic machine learning, with focus on multimodal and structured data. Structured data is any data that does not follow traditional iid assumptions, for example sequences, data on uniform grids (images, video), or data on graphs (protein networks, etc.). Traditional machine learning approaches cannot effectively model and do not scale to structured data domains. Mutlimodal data is characterized by data that simultaneously occurs (or has views) in different domains; for example, any social communication signal is often simultaneously conveyed through visually and aurally observed cues, which can be complementary, exclusive, or contradicting. Image data is frequently accompanied by textual descriptions or other types of auxiliary information, such as labels, etc. Traditional learning approaches cannot typically deal with this multitude of representations. I work on developing efficient models and algorithms, including deep Bayesian models (Variational Autoencoders), cross-domain and cross-modal adaptation & translation models, and fast meta-learning frameworks, to solve these problems.