Hardware to Enable A.I.
In my previous post I mentioned that I thought a major limiting factor to developing “real” A.I. will be hardware that supports massively parallel information processing. In short, my criticism was that existing A.I. solutions typically rely on brute force calculations and very powerful machines but that this was very different than what the brain does. I used the example of computing how a school of fish move by traditional modeling techniques versus how the real system does it, which is each fish figures out where to go next by itself and doesn’t need a master book-keeper tracking the movements of everyone and telling it where to go next.
I’ve been reading about IBM’s efforts to build hardware that is a step toward the type of information processing I had in mind. They call them Neurosynaptic Chips. I’m very much interested in learning how they write programs that utilize such functionality. I imagine this is a bit like trying to fit a square peg into a round hole. The brain’s “program” is not software, but instead is hard-coded by the structure of the connections between neurons. Or put another way, the “program” is in the hardware architecture itself. Perhaps a hardware/software model is more flexible than what the brain uses since one could conceivably use software to re-route information pathways instead of needing to alter the physical paths themselves. It’s far from obvious that nature’s solution to the intelligence question is the best one in principle, but we’ll have to wait until we understand how the brain does it before we can start parsing out the elements that are critical from the ones that are merely consequences of the constraints of evolutionary dynamics. It’s fun to think about!