Abstract: 

Recent advances in data-driven AI methods, generative models, and other techniques that can utilize large datasets have led to remarkable advances in generalization and capability. However, to create AI systems that can flexibly and resourcefully find novel solutions to new problems and obstacles they encounter in the real world, we need learning systems that can improve and adapt autonomously. Reinforcement learning offers a potential algorithmic framework to enable this. However, to make RL methods viable for real-world systems, we need to integrate the generalization capabilities that come from training on large prior datasets with the ability of RL methods to adapt on the fly. In this talk, I will describe how we can take steps toward making this possible, and discuss potential applications in robotics and other areas.

Bio: 

Sergey Levine received a BS and MS in Computer Science from Stanford University in 2009, and a Ph.D. in Computer Science from Stanford University in 2014. He joined the faculty of the Department of Electrical Engineering and Computer Sciences at UC Berkeley in fall 2016. His work focuses on machine learning for decision making and control, with an emphasis on deep learning and reinforcement learning algorithms. Applications of his work include autonomous robots and vehicles, as well as applications in other decision-making domains. His research includes developing algorithms for end-to-end training of deep neural network policies that combine perception and control, scalable algorithms for inverse reinforcement learning, deep reinforcement learning algorithms, and more.