On Jun 1, 2008, at 11:02 AM, Mark Waser wrote:
One is elegance. It would be "oh, so nice" to find one idea that would solve the entire problem. After all, everyone knows that the single concept of "neurons" is what our brains are built upon . . . . The problem is that they then take an incredibly simplistic view of what a neuron is and then can't figure out why they can't get it to work or why they have to use radically different simplifications and formulas to make it work in different circumstances.
Neurons *are* simple, analogous to a transistor. What they rarely seem to consider is how many different patterns and levels of pattern abstraction are required to make, say, a general purpose CPU design scale. You do not go from the 2,300 transistors of an Intel 4004 (nematode nervous system) to a modern CPU (reptilian nervous system) simply by slapping more transistors onto the 4004 design. Not only do you have to invent several new layers of abstraction, you also have to invent the control structures to manage all those abstractions and layers. All made out of simple transistors.
I think the general problem with neural networks is not the concept of the neuron but the notion that you can scale up the utility of a simple neural network simply by slapping more neurons onto it. It would be lovely if it was that simple, but I do not think the evidence supports the notion that the design can be both simple and efficient (in the sense that evolution would find a design to be "efficient").
J. Andrew Rogers ------------------------------------------- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244&id_secret=103754539-40ed26 Powered by Listbox: http://www.listbox.com
