I've always known the theory of NN's but never had the chance to get into the nutz and bolt of it. Not like Thomas or Eoin has been doing lately. So i thought i would post a general request for good text's on the topic. Hopefully one that can give me enough info to start developing them is software, not just a raw theory book.

So what is it? Whats your favorite NNet book if you have one....

Thanks :alright:
Posted on 2003-10-22 22:35:41 by NaN
The only NN book I've ever bought is Pattern Recognition by Self-Organizing Neural Networks. Its good but it conterates mainly on ART (Adaptive Resonance Theory) Nets.

To develop nets in software you really need to start with a problem and find the net which suits it best. From my study of nets they seem to all do the same general thing.

They take in a vector (array) of values, usually between 0-1 and output another vector. Generally speaking the output vector is treated either as

A) A category;

Say for example you inputted 20 vectors representing a plot of 20 functions, the net could categorise these into so many categories. You may have defined these categories beforehand (supervised learning) or not (unsupervised learning)

B) An input;

Say a neural net to solve the XOR problem, the output would be a single value from 0-1, or perhaps just 0 or 1. This value would be used as the input to something else. I think these pretty much have to be supervised learning thaught.

In Thomas' handwriting recognition the shape of the letter (or rather the stroke which created it) was converted in a vector, the pre-trained net then simply categorised these (I believe).

For a site with NN code and sample problems for each see Neural Networks at your Fingertips .
Posted on 2003-10-23 05:50:17 by Eóin
I already mentioned it somewhere on my site, but www.generation5.org is very good. Not a book though.. I have 'Artificial Intelligence' by George F. Luger because I needed it for my study, but it's still hard to understand after you've read it (the NN part that is).

Posted on 2003-10-24 17:12:49 by Thomas
"The Scientist and Engineer's Guide to Digital Signal Processing"
written in accessible way, complete,.... a must have.
Posted on 2003-10-24 23:36:29 by Ultrano
I've got an AI module this year, maybe my lecture's notes might be of some help.
Posted on 2003-10-25 08:24:03 by Eóin
Thanks all, this is fantanstic stuff. I would prefer buying a book since you can read it anywhere at anytime. However, the stuff being suggested is really good stuff to sink your teeth into!

If anyone else knows of a good book/link, please share it with us!
Posted on 2003-10-25 10:32:32 by NaN

I just want to add thanks for those notes! They are very good and easy to follow. I especially like the condensed feel to them. Most lecturer's like to go on and on beating a point into the ground that you forget where you were heading in general.

Again thanx!
Posted on 2003-10-26 13:21:19 by NaN
Posted on 2003-10-27 20:35:21 by bitRAKE
That too is an excellent source of info. That was the best telling i have read regarding those K. Maps, I now see (quite literally) what its point is. Well the self organization point anyway.

I still dont see how the end result will produce the solution to the travelling sales man problem? It never really explained the finished "learned" network regarding how it would then be applied and used.

If anyone wants to take a stab at it, im all ears ;)

Thanks alot bitRAKE for digging that one up!
Posted on 2003-10-27 22:09:07 by NaN
No digging - I'm still looking into creating a neural net in MMX - I want something like millions of nodes...

Goto the source (code):

Read about some applications:
Posted on 2003-10-27 23:33:59 by bitRAKE
I havent verified this but MIT has got its OpenCourseware on the web which has all its lectures online.
I guess they might have a section for AI too. You might want to check it out.

The link for MIT's openCourseWare site
Posted on 2003-10-28 08:29:35 by clippy
It might also be a good idea to refresh one's knowledge of the real neurons:
Posted on 2003-10-28 19:28:35 by bitRAKE
So does anyone know how to apply a leaned k-map to the solve the travelling sales man problem that it was bosted to solve?

The best guess i can think of is to say Im here (some how), and it will cause the closest node to have the highes value out of the map. Indicating where to go next. But i dont see how this will ensure you dont double back on youself (running in circles).

Posted on 2003-10-29 21:51:44 by NaN
I haven't bothered going through the Java code at the link I posted above - gone off to create my own neural nets designed for speed. They seem hard to train, so I am now trying to use a genetic algorithm to select net canidates.

The web site above said it solved a similar related problem.
Posted on 2003-10-30 00:22:41 by bitRAKE