Gregory Gundersen will present his General Exam Presentation on Monday, May 21, 2018 at 3:00 pm in CS301. Committee: Barbara Engelhardt (Adviser) Ryan Adams Mona Singh Title: Multimodal embeddings with structured priors Abstract: Medical images are visually evaluated by experts for key morphological features, but the genes underlying particular features are typically unknown. We propose a multimodal modeling and inference framework that learns modality-specific and interpretable latent representations of both genetic variation and medical image features. The method uses deep canonical correlation analysis (DCCA) to correlate paired embeddings. One modality's embeddings are constrained by a generative prior using a variational autoencoder (VAE). The full model is trained end-to-end with a loss function that combines the DCCA and VAE objectives. We demonstrate utility of this method by exploring the task of jointly learning point cloud clusters of histology slides and gene expression levels and recovering known biological knowledge. Barbara A. Mooring Interim Graduate Coordinator Computer Science Department Princeton University