Towards Training AI Agents with All Types of Experiences via a Single Algorithm
Speaker: Zhiting Hu, Carnegie Mellon University
Date: March 20, 2020, 2 p.m.
Host: Michael Overton
Training AI agents for complex problems, such as controllable content generation, requires integrating all sources of experiences (e.g. data, constraints, information from relevant tasks) in learning. Past decades of research has led to a multitude of learning algorithms for dealing with distinct experiences. However, the conventional approach to creating solutions based on such a bewildering marketplace of algorithms demands strong ML expertise and bespoke innovations. This talk will present an alternative approach from a unifying perspective. I will show that many of the popular algorithms in supervised learning, constraint-driven learning, reinforcement learning, etc, indeed share a common succinct formulation and can be reduced to a single algorithm that enables learning with different experiences in the same way. This allows us to create solutions by simply plugging arbitrary experiences in learning, and to systematically enable new learning capabilities by repurposing off-the-shelf algorithms.
Zhiting Hu is a Ph.D. student in the Machine Learning Department at CMU. He received his B.S. from Peking University. His research interests lie in the broad area of machine learning. His research was recognized with best demo nomination at ACL2019, best paper award at ICLR 2019 DRL workshop, outstanding paper award at ACL2016, and IBM Fellowship.
Refreshments will be offered starting 15 minutes prior to the scheduled start of the talk.