Re: virus: Re: neural nets

Alex Williams (
Sun, 5 Jan 1997 21:47:04 -0500 (EST)

> I would be interested in a summary of the Turing test discussion
> if you are willing. (Which mailing list?)

Recent development in a discussion on the Heavy Gear ML (dedicated to
discussing the setting/gameplay of the HEAVY GEAR RPG/tactical
wargame) of the feasability of a reinvading Terra using autonomous
`Bolo/Ogre'-like supertanks as a means of cutting down on the cost of
having to ship hujmans and support resources all the way from Earth.
Apparently, there are a lot of people out there that don't believe you
can build an autonomous combat agent given 3000yrs of development and
looked incredulous when I suggested given the Terra Novan NNet plug-in
modules they already have I could code up a subsumptive architecture
in 6mo that would be at least as good as a decently trained soldier
and not require major resource support.

The discussion of the modern limited Turing test emerged out of
demonstrations of present learning network and interaction system
discussions. Nothing fancy, I'm afraid most of the disputants know
about as much about AI as they do about police work ... from watching
PACIFIC BLUE, I'd imagine.

> There are two ways we can evaluate this. As an implementor of artifical
> neural networks I can assure you that computation is involved, after all
> I wrote the code that does the computations (i.e. summing activation
> levels over all inputs and mapping the result to an output function).

Of course, emulating a NNet at the level of software is going to
involve computation. That's the nature of emulation. See below. (I
don't envy you implimenting the poor beasts; I did it a couple times
and resolved to use someone else's software from then on if I possibly

> The other way is if the network is implemented in hardware or wetware,
> in which case the results are produced by virtue of electrical circuits
> rather than interpreting instructions. I would agree that there is no
> computation taking place *at that level* but when interpreted at the
> level of the network or higher, there may very well be computation.
> To illustrate, imagine a flipflop device implemented as a 4 neuron
> network (easily done), then using several of these flipflops to implement
> a NAND gate (also very straightforward), the using the NAND gates to
> build a computer.

True, one could do that, but the computation, at least in my opinion,
that the Von Neumann machine you /could/ construct from the NNet NAND
(and other logic gates) wouldn't be occuring at the NNet level any
more than a transistor is doing computation.

> Not if built in the way I just described.

But you're not emulating the whole system via NNet, you're emulating
componants and the level of functioning they produce performs
computation. (This is beginning to sound disgustingly like a
brain/mind distinction.)

> Allow me a question then: How is your "memesphere" different from
> a belief system?

A `belief system' is typically related to a single meme-complex; the
"memesphere" is made up from all the meme-complexi and stray memes you