Newsgroups: comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!europa.chnt.gtegsc.com!gatech!news.uoregon.edu!vixen.cso.uiuc.edu!usenet.ucs.indiana.edu!news.cs.indiana.edu!mmeiss@avocado.ucs.indiana.edu
From: "mark r meiss" <mmeiss@avocado.ucs.indiana.edu>
Subject: Re: Can neural nets think
Message-ID: <1995Jul1.173245.17879@news.cs.indiana.edu>
Organization: Computer Science, Indiana University
References: <3t2fmh$7cn@kaleka.seanet.com>
Date: Sat, 1 Jul 1995 17:32:38 -0500
Lines: 34

In article <3t2fmh$7cn@kaleka.seanet.com>,  <inficom@inficom.seanet.com> wrote:
>
>Well, "artificial", as in silicon electronic hardware, neurons have much faster switching and transmission speeds, but 
>you're overlooking the fact that real neurons have analog-valued outputs. In addition, synapses can have more than 
>one weight depending on the signal and where it came from. So, a real brain is much more powerful than even a very 
>large conventional ANN. My estimates are that in order to even be of the same order of magntiude of processing 
>power, you need to be able to perform quadrillions of (analog) connections per second. This is far above what we 
>can do today. But I think we will get there sometime fairly soon.
>

Several people have brought this "real neurons have analog outputs" point
up, but I don't really think it's relevant at all to the performance of
ANN's.  Using just 32 bits for a neuron's output, you already have over
four billion different output values possible.  You aren't going to 
convince me very easily that there are important values sitting in 
between some of those four billion outputs that give analog output an 
absolute advantage over digital output.  Or say there are--surely 2^64
possibilities is enough for you then?  The point is that digital outputs,
while always quantized, can approximate analog output to within any
detectable physical difference in the known universe.

There isn't any great hardware/software dichotomy inherent in neural 
networks that means that hardware system can do things of which software 
is fundamentally incapable.  In reference to functionality of universal
computing devices, hardware is just fast software.

Bottom line: you could simulate a human brain-sized neural network on
a Turing Machine if you had a realllllly long tape and you were
realllllly patient.


Mark Meiss (mmeiss@indiana.edu)
Indiana University CS Dept. Student

