Chapter
Controller Model Systems for Learning Predictive Models of the World
49:27 - 1:00:07 (10:39)
Controller Model Systems (CMS) provide the opportunity for the learning model of the world to predict potential outcomes using abstract high-level predictions, which are crucial for success.
Clips
The use of Long Short-Term Memory (LSTM) in reinforcement learning has been an improvement over traditional RNNs, but efficient planning among the many possible action sequences still needs to be addressed.
49:27 - 52:47 (03:19)
Summary
The use of Long Short-Term Memory (LSTM) in reinforcement learning has been an improvement over traditional RNNs, but efficient planning among the many possible action sequences still needs to be addressed.
ChapterController Model Systems for Learning Predictive Models of the World
EpisodeJuergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs
PodcastLex Fridman Podcast
In CM systems, controllers can learn how to use different parts of the model network to effectively solve new problems by reducing its search space.
52:48 - 56:47 (03:59)
Summary
In CM systems, controllers can learn how to use different parts of the model network to effectively solve new problems by reducing its search space. Even though the prediction is only for the next time step, the system can learn by memorizing important events that happened a million steps ago.
ChapterController Model Systems for Learning Predictive Models of the World
EpisodeJuergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs
PodcastLex Fridman Podcast
The use of physics simulations to teach behavior to machines for solving problems that humans can't solve has been a popular approach.
56:47 - 1:00:07 (03:20)
Summary
The use of physics simulations to teach behavior to machines for solving problems that humans can't solve has been a popular approach. However, developing a learning model of the world that captures abstract high level predictions is crucial for successful machine learning systems.