In this talk, Sergey Levine discusses how advances in offline reinforcement learning can help machine learning systems learn to make better decisions from data. (Video courtesy of CITRIS and Banatao Institute)
Wednesday, April 12th, Sergey LevineAssociate Professor of Electrical Engineering and Computer Science, Robotic AI & Learning (RAIL) Lab At the University of California, Berkeley, I gave the second of four Distinguished Lectures on the Current State and Future of AI. CITRIS research exchange and the Berkeley Artificial Intelligence Research Group (BAIR).
Examine Levine’s lecture Algorithmic advances that help machine learning systems remain both discriminatory and flexible. By training the machine using offline reinforcement learning (RL) methods, the machine can solve problems in new environments. Leverage large amounts of previously learned data and lessons while remaining adaptable to introduce new behaviors and thus new solutions.
As Levine explained, data-driven or generative AI techniques, such as the image generator DALL-E 2, can produce works that appear to be made by humans, but they can control robots to simulate humans. RL techniques, such as the Beating Algorithm, allow games to develop solutions that solve problems in unexpected ways. His research aims to discover how machine learning systems can adapt to unknown situations and make ideal decisions when faced with the full complexity of the real world.
Sergey Levine talks about using large datasets for reinforcement learning at the Center for Information Technology Research in the Interest of Society at UC Berkeley and the Banatao Institute (CITRIS).
“If you really want agents that are purposeful, purposeful, and able to come up with creative solutions, learning is not enough,” said Levine. “Learning is important, data is important, but the combination of learning and searching is a very powerful recipe.
“With unoptimized data, new problems cannot be solved in new ways. Optimization without data is difficult to apply in the real world outside of simulators,” he said. “If we can get both of these things, we might be able to get closer to this space exploration robot and actually come up with novel solutions to new and unforeseen problems.”
