Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11 - 2018

Details

Title : Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11 Author(s): Lex Fridman Link(s) : https://www.youtube.com/watch?v=3FIo6evmweo

Rough Notes

To Schmidhuber, it was clear that any machine should be able to learn how to solve problems and also improve the learning algorithm itself - he calls this meta-learning, and says that current meta-learning is more like transfer learning.

True meta-learning is about having the learning algorithm to introspection by the system using it, and be open to modification, from which the consequences of the modification can be evaluated and repeat the process over and over.

Universal problem solvers (e.g. Marcus Hutter's work) which exploit proof search brings an additive overhead, which cannot be ignored in many of the practical problems (which are small in the space of all problems) we want to solve in everyday life.

The history of science is the history of compression progress. We can use this to build artificial systems which measure the depth of the insights from data coming after experiments they perform, and giving an (instrinsic) reward proportional to this depth of insight.

PowerPlay searches for solutions to problems (e.g. path to get out of a maze) AND for 2 pairs of problems and their solutions where the system has the chance to phrase its own problem.

Creativity is a side-effect of problem-solving (i.e. search for the solution), rather than a module of its own. There are 2 types of creativity:

  • Applied creativity - coming from creative solutions to concrete problems.
  • Pure creativity - more like PowerPlay where the agent has the ability to choose their own problem to study.

Consciousness may also be a side-effect of problem solving.

Next wave of AI will be about systems that shape the data through their own actions.

Emacs 29.4 (Org mode 9.6.15)