From: J. R. Molloy (jr@shasta.com)
Date: Sat Sep 11 1999 - 04:04:20 MDT
>The problem isn't "wiring them together" but having them "do
>something" once wired. Nobody has ever had a 70 million cell
>neural net before. How deep do you make it? How broad do you
>make it? How many inputs & outputs should each neuron have?
Okay, I see what you mean. Just hooking them together doesn't accomplish
anything unless they can perform significant functions. (da-oh!) I had
skipped ahead to the concept of "self-assembly" which the HP guys did with
nano-wires, and applied that to the project of organizing de Garis' neural
net. With the proper encouragement, the neural nodes could connect
themselves in a similar way to how neurons in the brain grow connections on
demand. From what I can gather of the de Garis team's work, parallel
processing evolves algorithmically... so with self-assembling wires, just
let the machine decide how to wire itself up! Brains do it. Eco-systems do
it. Even marauding picnic ants do it. Let's let a massively parallel
processor do it.
Or... Perhaps I don't know enough about parallel processing to know that it
can't work.
>I'm not sure how flexible the de Garis architecture really is
>but if the depth can be up to 10 deep, and the net has
>1000 inputs and 100 outputs, and each neuron can interact with 500
>others, the possible number of configurations is huge. [I'm not an
>expert on neural nets, so if someone can explain this better
>please do so.]
Number of configurations (if limited to unweighted --purely digital --
values, i.e., |and|, |or|, |nand|, |nor| gates) without dialed-in
resistance, (since the wires self-assemble), and with allowance for
defective connections, can exceed that of the Internet. In addition, because
the connecting wires (analogous to dendrites) can grow in response to
directions from the processor nodes, analog computing results -- with all
the advantages of weighted signaling and attendant fuzzy logic.
My (laughably naive) intuition tells me that neural nets (or massively
parallel processors) and self-assembling wire want to converge with organic
computing (remember those flatworm neurons that learned how to do
arithmetic?) for some out of control AI.
Probably just a (double bowl) pipe dream. But the sloppiness of this
solution appeals to my lazy Buddha nature.
This archive was generated by hypermail 2.1.5 : Fri Nov 01 2002 - 15:05:06 MST