Efficient Interactive Learning - Learning More From Less

Explore efficient interactive learning algorithms for training decision-making agents like robots and large language models in this colloquium presentation by Gokul Swamy from Carnegie Mellon University. Discover how to overcome the limitations of static datasets by enabling agents to collect and learn from their own data, addressing two core challenges in reinforcement learning: exploration (experiencing the right things) and specification (knowing if experiences were good or bad). Learn about theoretical advancements and practical implications, including how to teach robots to provably recover from mistakes without extensive trial-and-error exploration, develop robust algorithms for training language models from conflicting preferences, and understand how reinforcement learning can achieve more with less data. Gain insights into promising directions for enabling the next generation of agents to learn more efficiently from limited data, moving beyond the era of simply scaling up training datasets toward more sophisticated learning paradigms that transcend internet-scraped human-generated data limitations.