Generally Intelligent Episode 28: Sergey Levine, UC Berkeley, on the bottlenecks to generalization in reinforcement learning, why simulation is doomed to succeed, and how to pick good research problems
17 snips
Mar 1, 2023 AI Snips
Chapters
Transcript
Episode notes
Levine's Shift to Machine Learning
- Sergey Levine switched from computer graphics to machine learning after realizing simulating minds was the key challenge in virtual humans.
- His first deep learning paper was in 2012 focused on deep reinforcement learning for 3D human motion.
End-to-End Deep RL Success Story
- Levine and his student John Schulman applied end-to-end deep reinforcement learning directly from pixels to motor efforts on a PR2 robot.
- Their experiments demonstrated end-to-end learning outperformed modular geometric pipelines in robotic manipulation.
Test Extreme Design Extremes
- In science, test how extreme a design can work to truly understand its value.
- Isolate new methods instead of patching with existing ideas to gauge their true utility.