Re: Web neurons

Fil Mackay (fil@mpce.mq.edu.au)
Fri, 24 May 1996 13:13:44 +-1000


Message-Id: <01BB4972.D6A29820@imedia.mpce.mq.edu.au>
From: Fil Mackay <fil@mpce.mq.edu.au>
To: "'John Middlemas'" <john@eco.powernet.co.uk>
Cc: "hyper-theory@math.byu.edu" <hyper-theory@math.byu.edu>,
Subject: RE: Web neurons
Date: Fri, 24 May 1996 13:13:44 +-1000

Hi John,

>It's sort of like a neural-network. If that is structured then so are Web
>neurons.

So, if a 'node' is a .html page, then how do concepts such as epocs (sp?), weights,
and transfer functions in neural nets map into your application?

Neural networks are optimisation (optimization for all you yankies :)) tools, I don't
see how you can create a popular medium out of it?

>What's a petri-net?

(I have only a vague recollection)

Basically a real simple network where you pass tokens between nodes. They
run on very simple rules (but do you think I can remember? either AND or OR).
I think you fire the first node, and the changes cascade through the network.

PLEASE - anyone who knows exactly what a petri-net is, please describe it, and
stop my dribbling!

>>>I don't know much about Java but I don't believe it operates within a Web
>>>page which is the important point (I haven't got Windows 95 or NT).

>>Uhh.. Java operates within web-pages (applets) and as stand-alone applications
>>(in their own window). What environment do you run on?

>poor old Windows 3.1  

Win 3.1 runs Netscape 3.0 beta which runs Java (from memory).

Developing under 3.1 is difficult, because of the short file-names. No quick
solution to that IMHO.

>Can you view java as lots of small Web page tables and trace the logic as I
>mentioned above? or is it gobbledegook as usual?

Uhh.. more information? You can run several applets (which could be like sub-pages)
on the one 'page', yes.

Trace the logic.. you mean as in a debugger?

>>>Why complicate by adding another language when all that is needed is some
>>>modification to HTML. Integration is better than fragmentation.

>>Sometimes it's better to wipe the slate clean and use something better.

>Will Java do away with HTML then?

Good question. Perhaps, but not for the short term. When Java gets integrated
into every OS under the sun, there'll be a good reason. Perhaps what we call
'browsers' will become layers of Java (like HotJava), which will open the web
up a lot. HTML will then have to compete with new formats.

>How will pages be rendered?

My guess is that Java code will render the pages.

>I don't believe Java has any neural basis. I would love to start from
>scratch but is there time?

How long do you have? :)

Java does not have a neural basis, though implementing a neural application
in Java would be no problem-o.

>I don't follow this. I checked my .htm file sizes and some are as low as a
>few humnred bytes?

File systems need to store files in chunks of 'x' bytes. My FAT file system is 64k
chunks (pretty large), so a file of 1 byte actually consumes 64k of physical space.

I don't know too much about UNIX filesystems, but your files (on your FAT) system
would probably be taking 16-32k of physical space. Have a look at the cluster
size using chkdsk.

>>>>Why not use a neural net model to represent content/the web? Then you could
>>>>build all these structures?
>>
>>>I don't know how to.
>>
>>Why not use a simple table:
>>
>>ID, Description, FromID, ToID
>>
>>This table is able to model a neural network. Each entry is a node in the
>>network, and
>>can potentially be a link (using FromID, ToID).
>
>Now we're getting somewhere.

What I said is incorrect, I think. Well, it's not exactly a 'neural net', although it could
be. I'd call it a simple symbolic network. It could be used to describe anything (eg.
A neural network).

>This is the sort of thing I am trying at the
>assembler level although didn't know it was a neural-net model. But couldn't
>we also use HTML tables - perhaps one per page?  

IMHO you are going to need more nodes than you realise.

>How would you implement your idea to produce say the sort of distributed
>chain reaction I mentioned above and also keep normal surfing ability, which
>people won't want to lose unless there's a replacement.

What exactly do you mean by 'distributed chain reaction'?

>Should there be some input and output levels as well:
>
>ID,   Description,   (FromID, input_level)     (ToID, output_level)
>
>After all each neuron in the brain has inputs and outputs that pulse at
>variable frequencies. 

I don't understand what these weights (levels) are going to do? Under a classic
optimisation problem (a normal n.net application) I/O is quantitative. How is
this so with surfing the Web?

Later..