Wednesday, October 22, 2014

old ideas, restated

an argument. intelligence is a vast collection of well-interfunctioning solutions accumulated over biological history. if you want to recreate something like intelligence, you have to figure out many of these kinds of solutions and put them together in a meaningful way.

for example, the mechanical shapes of axon guidance proteins for visual cortex are probably tuned in some way to natural visual statistics, or to meta-aspects of the visual problem. even the shape of the body itself (like having legs and stuff) is a kind of intelligence. there's an incredible array of clever engineering tricks throughout all levels.

a counterargument. maybe there are simpler principles underlying how these solutions get found and put together. so if you had a grasp of these principles, you wouldn't have to spend decades researching solutions, but instead you could turn your principles loose on tons of real data and they would gradually figure out all these solutions similar to how real life did it.

that's why i was excited about "plasticity in overlapping representations". if there are simpler principles, i think they'd have to erase the boundary between form and content in a learning system.