Juergen Schmidhuber on Gödel Machines, Meta-Learning, LSTMs, and the Simplicity of Universal Intelligence

Juergen Schmidhuber discusses his lifelong ambition to create AI systems capable of recursive self-improvement, a dream born from a desire to amplify human creativity to solve the universe's mysteries. He introduces "meta-learning" as the core mechanism, where a learning algorithm can inspect, modify, and improve itself, leading to a hierarchy of self-improving systems. This concept is contrasted with "transfer learning," a more common practice in deep neural networks today, which involves reusing pre-trained layers for new tasks rather than fundamentally altering the learning process itself. Schmidhuber also delves into "Gödel machines" and universal problem solvers, which are theoretically optimal but currently impractical due to significant computational overheads associated with proof search.

A key distinction is drawn between these asymptotically optimal universal problem solvers and the more practical, non-universal methods like recurrent neural networks trained by gradient descent. While universal solvers guarantee optimality for large problems, their constant overhead makes them unsuitable for the "small" problems (which can still be computationally vast for humans) that current AI excels at. Schmidhuber argues that the most successful practical AI solutions, such as LSTMs, are often surprisingly simple at their core, built upon millennia of mathematical abstractions. He believes that a truly general intelligent system will also possess this underlying simplicity, expressible in just a few lines of pseudocode.

He advocates for continued theoretical research to bridge the gap between these two paradigms, aiming for a theory that accounts for limited resources in our universe to achieve practically optimal problem-solving. Schmidhuber extends his ideas to the philosophical realm, proposing a deterministic universe where quantum randomness is merely pseudo-randomness, arguing that simplicity and compressibility are hallmarks of beauty and elegance in the universe's fundamental laws. He views the history of science as a "history of compression progress," where new theories simplify and unify observations, leading to deeper insights and "scientific fun."

This concept of compression and the pursuit of simple explanations underpins his ideas of artificial curiosity and "power play." In "power play," AI agents are not just problem-solvers but also problem-generators, intrinsically motivated to discover and solve the simplest new problems that expand their generalization abilities. This mirrors human curiosity, where individuals and even babies act like scientists, exploring the unknown to build a more general understanding of the world. Schmidhuber suggests that evolution has instilled this curiosity strategy in humans as a means of survival and exploration, highlighting its essential role in both human and artificial intelligence.

I thought the most exciting thing is to solve the riddles of the universe and and that means you have to become a physicist however then I realized that there's something even grander you can try to build a machine that isn't really a machine any longer that learns to become a much better physicist than I could ever hope to be.

true mental learning is about having the learning algorithm itself open to introspection by the system that is using it and also open to modification such that the learning system has an opportunity to modify any part of the learning algorithm and then evaluate the consequences of that modification and then learn from that to create a better learning algorithm and so on recursively.

The universal problem solvers like the Gödel machine but also Markos who does fastest way of solving all possible problems which he developed around 2012 - in my lab they are associated with these constant overheads for proof search which guarantee is that the thing that you're doing is optimal.

there's nothing more practical than a good theory and um yeah and a good theory of problem-solving under limited resources like here in this universe or on this little planet has to take into account these limited resources.

experience tells us that this stuff that works best is really simple so see asymptotic team optimal ways of solving problems if you look at them and just a few lines of code it's really true although they are these amazing properties just a few lines of code.

as long as we don't have evidence that there is no short program that computes the entire history of the entire universe we are a scientists compelled to look further for that Swiss program.

indeed the history of science is a history of compression progress.

power play just goes one little step further and says let's not only search for solutions to a given problem but let's search two pairs of problems and their solutions where the system itself has the opportunity to phrase its own problem.