Modeling interactions with Deep Learning

Interacting systems are highly prevalent in many real world settings, including genomics, proteomics, and images. The dynamics of complex systems are often explained as a composition of entities and their interaction graphs. In this dissertation, we design state-of-the-art deep neural networks for interaction-oriented representation learning. Learning such structural representations from data can provide highly accurate predictive models, semantic clarity, and ease of reasoning for generating new knowledge. We consider three different types of interaction graphs: 1) interactions within a particular input sample from a functional genomics task, 2) interactions between multiple input samples from a proteomics task, and 3) interactions between output labels from a computer vision task. For each type of interaction, we design novel models to tackle a real world problem and validate our results both quantitatively and visually. We show that deep learning models with relational biases can learn representations of entities and their interactions, as well as enable us to discover the governing dynamics.