Scalable and Flexible Distributed Reinforcement Learning Systems

Explore the intricacies of scalable and flexible distributed reinforcement learning systems in this 51-minute lecture by Bo Zhao at the Finnish Center for Artificial Intelligence FCAI. Delve into the challenges of building efficient machine learning systems that translate data into valuable decision-making tools. Examine recent breakthroughs in large language models and reinforcement learning, emphasizing the importance of scalable model training on large GPU/TPU clusters. Learn how to co-design multiple layers of the software/system stack to enhance scalability and performance of ML systems. Discover techniques for creating flexible distributed RL systems that accelerate and parallelize the RL training loop, as well as statement management libraries for transparent GPU device allocation and multi-dimensional parallelism. Gain insights into open challenges in designing large-scale RLHF systems from Bo Zhao, an Assistant Professor at Aalto University specializing in efficient data-intensive systems and compilation-based optimization techniques.

Bo Zhao: Scalable and Flexible Distributed Reinforcement Learning Systems