Hyperparameter Dynamics
Tom discusses the impact of adding a new hyperparameter to the architecture and how it affects the learning dynamics in different games. Tim explores the idea of using neural networks as proxies for non-differentiable parameters in meta learning, highlighting the complexity it introduces to the system.In this clip
From this podcast
Machine Learning Street Talk (MLST)
#49 - Meta-Gradients in RL - Dr. Tom Zahavy (DeepMind)
Related Questions
Are there similar strategies for learning different skills as discussed in the episode Episode 30: Ben Eysenbach, CMU, on designing simpler and more principled RL algorithms and the clip Skill Learning Insights?
Is meta-learning a form of intelligence as discussed in the episode Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10 and the clip Hierarchical Learning Insights?