Lex Fridman Podcast Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs
21 snips
Dec 23, 2018 AI Snips
Chapters
Transcript
Episode notes
Metalearning vs. Transfer Learning
- True metalearning lets systems introspect and modify their own learning algorithms.
- This differs from transfer learning, which reuses existing knowledge for new tasks.
Gödel Machines and Practicality
- Universal problem solvers like Gödel machines have a theoretical, asymptotic optimality.
- A constant overhead makes them impractical for smaller, real-world problems.
Simplicity of AGI
- A general intelligence system might eventually have a simple, elegant underlying algorithm.
- This simplicity builds upon millennia of abstractions, like mathematics and calculus.