Relational Structure Discovery for Deep Learning

Structure is ubiquitous: from fundamental physical interactions to molecules, biological interactions, social networks and many more spread across the universe. Not only is the world around us rich in structure, but our mental model of the world is also structured: we think,
reason, and communicate in terms of entities and their relations. Such a graph-structured real world calls for artificial intelligence methods that think like humans and hence employ this structure for decision making. Realizing such a framework requires known structure/graph as well as models that can ingest these non-linear graphical inputs. In cases of a latent unknown graph structure, state-of-the-art deep learning models either focus on task-agnostic statistical dependency learning or diverge from such explicit feature dependencies during prediction. We bridge this gap and introduce methods for jointly learning and incorporating graph based relational knowledge into state-of-the-art deep learning models to help improve (1) predictions, (2) interpretability, (3) post-hoc interpretations, and (4) test dataset selection. Specifically, we contribute methods that enable learning graphical relationships from data in the absence of a ground truth graph. Furthermore, we introduce plug-and-play methods that bias deep learning models to include this graph explicitly for improving the aforementioned stages of a deep learning pipeline. We demonstrate these capabilities on simulation, tabular, text classification and vision tasks.