1 post tagged “ai”

Last week on the TWiT's Futures in Biotech episode 10 Marc Pelletier interviews Dr. Carla Shatz about the human brain. The most fascinating part of the entire interview was a brief part discussing how the brain manages to connect the retina to the visual cortex without ending up completely random. This wasn't the first time I had read about the mechanism – V. S. Ramachandran has at least mentioned it in one of his books – but I had pretty much forgotten.

What happens is that the retina projects nerves to the back of the brain to connect with the occipital lobes to connect into the primary visual cortex. The connects are for the most part random, which is of course bad  because
there's no sense to the signals. Waves of activation spread across the retina, neighboring neurons lighting up with electrical activity while more distant ones remain quiet, sending signals back to the visual cortex. The visual cortex then is able to coordinate the input, since even though the receiving neurons seem to randomly light up, they only light up because they neighbor one another in the retina. I suspect it uses some form of Hebbian learning, where neurons that fire simultaneously tend more often to be affected by one another's activation and those that don't are less affected.

To illustrate I offer a simple example using a string of characters. Suppose you have a string that's been scrambled, say "     bbeenoooorttt". With information about neighbor status (the first t is next to the start of the string and next to the first o, the r is next to the second o and the first space, the second space is next to second and third t's, etc.) you can reconstruct the original: "To be or not to be". By using the neighbor status of the letters you can very simply say what they are near and thus by knowing what everything is near, you know where everything is.

there are a number of interesting applications that this might have, such as perhaps self building microchips, but the one I'm particularly interested in is artificial neural networks. Self organizing properties would make it significantly easier to design neural networks because it would remove the necessity to design the thing down to the tiniest detail. Instead they could be designed in a broader sense, by function rather than specific wiring. All that's left to do is get cybernetic implants to work properly and we'll have working cyberbrains! But that's another project...