Relational Structure Discovery for Deep Learning
Abstract:
Cognitive psychologists have identified relational structure as one primary means by which humans tackle unstructured problems. Our primary relational thinking strategy simplifies complex systems as compositions of entities and their interaction graphs. We borrow such an idea and design graph-oriented relational representation learning into state-of-the-art deep neural networks. Existing deep learning literature has proposed effective ways, like using graph neural networks, to represent data when relational graphs are known a priori. However, little attention has been paid to address cases when the underlying relation graph is unknown. We propose 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 evaluation.
Committee:
- Matthew Dwyer, Committee Chair, (CS/SEAS/UVA)
- Yanjun Qi, Advisor, (CS/SEAS/UVA)
- Yangfeng Ji (CS/SEAS/UVA)
- Vicente Ordóñez Román (CS/SEAS/UVA)
- Jianhui Zhou (Department of Statistics/GSAS/UVA)