Chapter
Neural Networks and Transfer Learning
13:07 - 23:12 (10:04)
This episode discusses the concept of neural networks and how they can tile the space, making deep mind impressive results possible through transfer learning even in the real world.
Clips
The use of simple control architecture with linear feedback controllers can do a lot to stabilize complex dynamic systems like a helicopter in stationary flight, although the real challenge comes with the difference in time scales in the real world.
13:07 - 17:05 (03:58)
Summary
The use of simple control architecture with linear feedback controllers can do a lot to stabilize complex dynamic systems like a helicopter in stationary flight, although the real challenge comes with the difference in time scales in the real world. Neural networks, on the other hand, learn to tile the space, enabling them to do more complex tasks, compared to linear controllers or finite state machines.
ChapterNeural Networks and Transfer Learning
EpisodePieter Abbeel: Deep Reinforcement Learning
PodcastLex Fridman Podcast
The combination of deep learning processing with traditional underlying dynamical systems for planning has been challenging, but it can be achieved by choosing a latent variable that informs about the future before taking high level action, which leads to faster learning with better credit assignment.
17:05 - 19:15 (02:10)
Summary
The combination of deep learning processing with traditional underlying dynamical systems for planning has been challenging, but it can be achieved by choosing a latent variable that informs about the future before taking high level action, which leads to faster learning with better credit assignment.
ChapterNeural Networks and Transfer Learning
EpisodePieter Abbeel: Deep Reinforcement Learning
PodcastLex Fridman Podcast
The challenge of AI lies in transfer learning, generally between learned models.
19:15 - 23:12 (03:56)
Summary
The challenge of AI lies in transfer learning, generally between learned models. Through meta-learning concepts, this issue is addressed, and big models trained on multiple objectives could get extended to real scenarios.