Chapter
Combining Rule-Based Systems with Neural Networks for Better Generalization
1:12:56 - 1:26:01 (13:04)
The speaker discusses the benefits of combining rule-based systems with neural networks, specifically in the context of StarCraft, where expert systems and discrete decision-making can improve the performance of bots built with neural networks. He also emphasizes the importance of sequence to sequence learning in allowing neural networks to learn a function that produces any output.
Clips
Discrete decision making in neural networks can offer better guarantees of generalization, particularly relating to out-of-distribution data.
1:12:56 - 1:17:08 (04:11)
Summary
Discrete decision making in neural networks can offer better guarantees of generalization, particularly relating to out-of-distribution data. While increasing the size of the training set is one solution, incorporating a programmatic decision-making approach in addition to neural networks can further improve performance.
ChapterCombining Rule-Based Systems with Neural Networks for Better Generalization
EpisodeOriol Vinyals: DeepMind AlphaStar, StarCraft, Language, and Sequences
PodcastLex Fridman Podcast
State machines and rule-based systems are combined with neural networks to make the system the best that detects cancer or weather patterns using the magic of machine learning.
1:17:08 - 1:21:39 (04:30)
Summary
State machines and rule-based systems are combined with neural networks to make the system the best that detects cancer or weather patterns using the magic of machine learning. Sequence-to-sequence learning allows input of anything to a neural network for learning a function that produces any output, including image captioning and text translation.
ChapterCombining Rule-Based Systems with Neural Networks for Better Generalization
EpisodeOriol Vinyals: DeepMind AlphaStar, StarCraft, Language, and Sequences
PodcastLex Fridman Podcast
Self-play can quickly develop reasonable strategies in StarCraft despite long sequences of actions in games.
1:21:39 - 1:26:01 (04:22)
Summary
Self-play can quickly develop reasonable strategies in StarCraft despite long sequences of actions in games. This is possible with at least 10 minutes per game on average, making it quicker than relying on supervised learning for large LSTMs or transformers.