Contextualized embeddings have revolutionized the field of machine learning. First, as a means to encode text in natural language applications and later on as a representational mechanism for other modalities including image, longitudinal, and high-dimensional structured data. In recent years, embedding approaches have been proposed to address problems in biology and healthcare, however, there are many important questions that require further investigation. For instance, i) how to effectively integrate embeddings for discrete elements with continuous measurements, ii) how to integrate granular temporal or ordering information into embeddings, and iii) how to effectively create embeddings for multimodal data. Successful applicants will work toward developing a model prototype addressing one of the questions above using state-of-the-art representation learning approaches based on deep learning architectures.