Newsgroups: comp.ai,comp.ai.neural-nets,comp.ai.nat-lang
Path: cantaloupe.srv.cs.cmu.edu!bb3.andrew.cmu.edu!newsfeed.pitt.edu!godot.cc.duq.edu!newsgate.duke.edu!news.mathworks.com!newsfeed.internetmci.com!swrinde!howland.reston.ans.net!ixnews1.ix.netcom.com!netcom.com!i17
From: i17@netcom.com (Valucard International)
Subject: Re: * how many neurons for language output?
Message-ID: <i17Du2oL2.AD9@netcom.com>
Followup-To: comp.ai,comp.ai.neural-nets,comp.ai.nat-lang
Organization: NETCOM On-line Communication Services (408 261-4700 guest)
X-Newsreader: TIN [version 1.2 PL1]
References: <4r8tir$2s0k@lamar.ColoState.EDU>
Date: Fri, 5 Jul 1996 13:39:01 GMT
Lines: 47
Sender: i17@netcom17.netcom.com
Xref: glinda.oz.cs.cmu.edu comp.ai:39770 comp.ai.neural-nets:32390 comp.ai.nat-lang:5179


Wow.  

Could someone please say what the difficulties are with using a
NN of several layers to look at (let's say text for now but audio is 
just a step farther) speech or dialogues and gradually be able to "take
over" or predict the flow of text?

Instead of telling me that this won't work, I'd rather hear about what has
been tried along these lines.  How big a net was used?  How slow was it?

A completely ad-hoc system with no preformed dictionaries or "concepts",
just pure pattern-flying.  Sort of like a body-snatcher, or a human baby,
listening and babbling until it asymptotically "makes sense".  

Are there plateau at which such systems "click in" and start looking pretty 
good?  Of course there are.  I operate like this all the time, and I've 
seen people with much less in the way of auxiliary concepts and theories
than me, who do a marvelous job functioning in this world.  


17


Steven Eisenberg (eisenber@lamar.ColoState.EDU) wrote:
: How do you determine how many neurons to use for both the input layers 
: and output layers for language output?  For example, if I want to input a 
: sentence into a network, how do I determine how many neurons would work 
: best for the input layer?  Is it best to use one neuron for each word, or 
: break down the word into letters and put each letter into a neuron?  How 
: would you assign a value to each word as input into the network?

: For the output, would there be one output neuron for each word, or would 
: one neuron suffice and print (or speak) one word at a time?  Which gives 
: best results?

: Thanks.

: Steve Eisenberg



-- 

------------------------------------------------------------------------------

                      May the best hallucination win.


          I want a God who takes responsibility for His mistakes.

------------------------------------------------------------------------------
