Can Curiosity Be Programmed? 269
destinyland writes "AI researcher Jurgen Schmidhuber says his main scientific ambition 'is to build an optimal scientist, then retire.' The Cognitive Robotics professor has worked on problems including artificial ants and even robots that are taught how to tie shoelaces using reinforcement learning, but he believes algorithms can be written that allow the programming of curiosity itself. 'Curiosity is the desire to create or discover more non-random, non-arbitrary, regular data that is novel and surprising...' He's already created art using algorithmic information theory, and can describe the simple algorithmic principle that underlies subjective beauty, creativity, and curiosity itself. And he ultimately addresses the possibility that the entire Universe, including everyone in it, is in principle computable by a completely deterministic computer program."
programming (Score:3, Insightful)
I think that the approach commonly taken to achieve some form of AI (curiosity as an example) through programming methods may be a flawed way of going about it. We probably should go about the problem in a similar way to how biological systems developed various aspects of AI. That is, build a system that has some basic rules for its operation that tends to form a system where curiosity and intelligence in general is an emergent property rather than one that is strictly programmed into the system. Take an existing system with some degree of "creativity" inherent in it and model our own technology to at first, mimic the natural system and over time, we tweak the system to suit our purposes as It is extremely difficult to build such systems from scratch.
AI has fully matured (Score:1, Insightful)
I'm glad to see serious researchers are at work figuring this stuff out, now that they've got a working definition of intelligence and have figured out how to make intelligent programs.
Show me the runny (Score:4, Insightful)
Schmidhuber has interesting claims, like about his Goedel machine [idsia.ch], an algorithm that makes provably globally optimal self-modifications.
But he never seems to get around to actually writing the code, or even non-vague pseudocode to implement these algorithms to show how they actually work and that they actually work. I guess it's just an "implementation issue". Ah, the chorus of the pure theorist...
Re:What's that like? I'm curious. (Score:5, Insightful)
Depends greatly on what you research. Unfortunately, the vast majority isn't as glamorous as you might imagine. I work in a pretty interesting area (an academic area with connections to videogame AI and game design), and this sort of creativity / discovery-systems / curiosity / art / etc. research is interesting too. But the vast majority is more pedestrian. Sure, there's interesting applications: computer vision, robotics, planning, data mining, bioinformatics, etc. But 90% of the work that comes out is incrementalist stuff; relatively boring proofs of some fact, or new algorithm that's 7% faster in some important special case (I suppose that's true of a lot of scientific fields, though).
It goes back and forth in waves, though. It seems that there will be waves of pretty exciting AI research, then a backlash as some of it goes over the top into sci-fi Singularity Is Nigh sort of AI, then things swing all the way to the other direction into AI as a really narrow field that's basically applied statistics, control theory, symbolic logic, and planning, and the only stuff that can get published is Rigorous stuff with Proofs (sort of a defensive reaction by people worried about being branded kooks). Then after a few years of that everyone realizes that 5000 more proofs in some super-narrow area aren't getting us anywhere because the field is stagnant with no direction, and people start doing more speculative applications and proposing new problems again. Then repeat.
It's somewhat unfortunate on the whole that there's such a big gap between what you might call "layperson AI" and "academic AI". The layperson AI (the singularity crowd, etc.) are excited about stuff, and have interesting goals, etc., but often do stuff that verges more on the sci-fi than the scientific. But academic AI is so scared of being them that it consciously tries at times to be super-boring so nobody mistakes them for Hans Moravec.
Re:Physics of computing the universe (Score:3, Insightful)
Re:Non-determinism. (Score:5, Insightful)
No more ridiculous than the idea that Consciousness is reducible to basic, known, physical laws.
Re:Physics of computing the universe (Score:1, Insightful)
> We should, in theory, be able to simulate the universe, just not as fast as the universe actually moves.
No, you couldn't even simulate the entire thing slower, because you'd need to be able to store the full state of every particle in the universe. Due to quantum uncertainty, you couldn't even measure the full state of any particle without changing it, but even if you could, you couldn't store more than one particle's worth of data in less than one particle. So it would take a computer the size of the universe just to store the data of the universe. And that computer would have to count itself as not part of the universe, otherwise we have an infinite storage loop.
Note: you could simulate *a* universe (just do a much smaller one, so that you could actually store it), just not *our* universe. And simulating *a* universe would still highly scientifically useful if it ran on the same rules and ran fast enough to test things out.