The title of the project is: “An Integrated Nonparametric Bayesian and Deep Neural Network Framework for Biologically-Inspired Lifelong Learning” with MIT and UC Berkeley.
There have been remarkable advances in recent years using deep neural networks, but generalizing them to new situations has proven significantly more difficult, presenting barriers to the creation of many of the artificial-intelligence-based systems. In order to create systems that are capable of learning continually, generalizing to novel situations, and applying prior knowledge when appropriate, we look to incorporate Bayesian modeling into the successful framework of deep neural networks. Interestingly, the limitations of deep neural networks are strengths of Bayesian modeling, and we aim to create a novel integrated framework, the Infinite Task Network (ITN), that unites the best of both to overcome the limitations of using either in isolation.
The award is from DARPA, the lifelong learning (L2M) program.
Award amount is $3,629,602. I'm the PI. The Co-PIs are Cynthia Rudin and Larry Carin (at Duke), Micheal Jordan and Tom Griffiths (at Berkeley) and Josh Tenenbaum and Tamara Broderick (at MIT).
Statistical Science
