Saturday, March 07, 2009
when jeff hawkins spoke at sfn, he made the point that humans receive a lot of training before they can do anything intelligent. there's just so much stuff in the world that you have to be exposed to a lot before you can start to sort it out. i looked at the numenta website just now and i saw that they still haven't made any significant progress on AI. their program can recognize a sailboat, as long as the sail is up and it's oriented the right way and so forth. then i thought, to recognize other configurations, the program would have to have some deep knowledge about sailboats and how they work, not just pixels. so it would probably have to be trained on other kinds of data besides pictures of sailboats. then i wondered if this could actually be an advantage in some way, because of the repeating nature and scale-invariance of patterns in the universe. for example, you can attach a mast to a sailboat like you attach a methyl group to a carbon ring.