Status : PhD Student
Title of thesis : Understanding Deep Neural Networks Through Their Internal Representations
Research Topic / Interests : Generalization in DNNs has long been theoretically misunderstood. Through studying the vectors of activations in a layer as they are seen by neurons in the network, we hope to find what characterizes the internal representations in networks with good generalization. We also briefly cover how we hope to use this knowledge to theoretically underpin known best practices in DNN training, and give some potential industrial impacts of our research.
ARIAC Work Package : WP2 - trust mechanisms for AI : federated learning & block chain, inductive logic and predictive justice, formal methods & certified AI , robust IA, distillation, infovis, digital assistants