Research Project Title:
Graph Evolution
abstract:Graph networks have recently shown promise in providing an abstract tool to capture relational information between objects using neural networks. The naive assumption of a fully connected graph offers little in the way of inductive bias to facilitate learning. As a result, designers often resort to hand-crafted relations, such as K-nearest neighbors. Besides creating additional hyperparameters to tune, these models are insufficient for general object reasoning and preclude utilizing any information in the scene to prune relations in a deliberate manner. We hypothesize that discrete optimization techniques such as evolutionary algorithms may provide a means to constructing optimal relational patterns for graph input into neural networks.
About:
"I hope that through this SuperUROP I will learn how to write a publication for an academic conference. Before this I've had multiple research opportunities in machine learning and the application of neural networks, but none have led to a tangible publication. I am excited to explore a more novel area of deep learning and work towards tangible products that demonstrate my research experience and prepares me for graduate level study."
