Calling SAM FINCH

[Gen-X]

Hi Sam

I noticed in one of your other posts that you are heavy into Mathematics and Artificial Intelligence.

I was wondering then if you could answer a few questions that have been plagueing me :

I have done some reasearch into AI, specifically Neural Networks and while all of my information to date tells me everything I need to know about how it is done, feedback looks, nodes, triggers and everything else...

It lacks one VERY specific thing...


How do you decide how the network is set up?

I mean it describes in quite complicated mathematics about the formulae for each node and all that but it doesn't relate to a real world environment.

Lets say for instance I want to use a Neural Net to try and break down a sentence.

"Hello, my name is John"

How would I use a neural net to do this?
What would be the nodes?
What would be their formulae?
How do I inter connect them?


If you could provide even the slightest inkling for some of these would be muchly appreciated.

Thanks

[Vuen]

uh huh... maybe I will manage to understand when I become 16 too... but for now its just simple .AF hacking...

[Paul282]

If you're interested in Language processing you should take a look at a language called Prolog (programming logic). Not so much to use as a replacement for VB but as a different way of thinking about the toppic. Prolog is relation based so it's syntax is designed around relations of people and objects and rules.

EG...

Says(John, "Hello, my name is John" )
Green(apple)
Likes(John, Mary)

plus rules etc. It's a little different and sounds a little lame here but you should take a look at it there are tutorials all over the net.

There is also a topic on 'Real Language Processing' for extracting meanings via noun and verb phrases but I can't remember what it is called

I'd like to hear the answer to your question about Artificial Nerual Networks too though

[Sam Finch]

There are 2 types of artificial intelligence, there's the one I can to and the one I can't do and you wan't to.

The sort of AI I do is where the computer's dealing with a specific proplem and you can break down the mathematics so you can give the computer an algorythm that it can follow to get the solution and appear intellegent.

The other sort is using nural net's to make the computer actually intelligent (do you remember that creatures program some time ago?) this is done by setting up a nural network and some binary data that it can set itself up with and never raise an error(ie an array representing connections between nodes) these bits of binary data are tested to see which one does the best and then the best ones are combined to make new ones, this process of evolution is then repeated until something fairly intelligent comes out the other end.

I think what your trying for is something in between, This is not my area of expertise but I would like to get into it.
If you can reccomend any books, websites etc where I can learn I would be verty interested and whould post anything I found that would be helpfull, other than that I can recommend 2 areas of mathematics, Graph theory and Computability theory, I think most of your answers would lie in there. Group theory and ring theory would help too, They're all quite nice branhces of maths in that you don't need to no much other maths to learn them. (you need to know group theory to learn ring theory and you need some basic maths to do computability theory but other than that you're ok) I you point me towards some AI basics I would be happy to do some research for you.

[Gen-X]

Hi Sam,

Thanks for your reply.... I hear what your saying and its a branch of Artificial Intelligence called "Genetic Algorithms"... and I have a fairly good understanding enough on that... But it does suffer from being a glorified "Search the best for me" rather than actually mimicing any form of intelligence.

I am not exactly after something in between... but I guess you could say that I am after a non-mathematical form of a Neural Network.

I suppose it would be possible to intersplice a set of boolean characteristics into a latice nodule structure such that you could provide a "rough" statistical probability for performing a certain task based on the level of input acquired.... but this would again be very specific to the task at hand.

What I am thinking is that perhaps I would need to prioritize the nodes into 3 seperate groups...

The first tier : (INPUT nodes)
These would be nodes that are triggered should the specific input be given. The formulae for these could be quite easy as it would be either a boolean (Is what I am looking at alive or dead?) or a simple formulae (Is what I am looking at LARGE? If their height / my height > 1 then YES).

The middle tier : (PROCESS nodes)
These would be what you could consider as "thoughts" which describe attributes or inclinations towards certain dispositions as a result of the triggered neurons from the first tier. Hopefully once several of these have been triggered and interact with others at least "ONE" of the final tier is triggered.

The last tier : (ACTION nodes)
Each of these nodes would result in a specific "action" that could be taken as a result of the processed inputs. Ideally it would involve some form of backwards stepping and private determination as to what needs to happen in order for this to be triggered.

I see the main problem occuring in defining the loopbacks but then again it could always be used to adjust the "threshholds" that were used to make wrong or right decisions...


Mmmm... I may have to think about it some more.

[Justin Ward]

Dudes,

As a geophysicist and mathematician, I use and recommend Neurosolutions V3.02. Check out www.neurosolutions.com.
I'm not on commission, I just realy like the (demo) program.

You can download a demo version from the site that includes demos. The site covers grammer applications, and neural network architecture. It uses a highly graphical interface that attracts new users and yet is powerful enough to keep experts interested. All networks can be output as VB or C dll's. Wizards help you generate NN's from scratch.
Genetica algorithms are now supported and there is an online book for beginners.

Sam, as mentioned by Paul282, Prolog is probably the language you are after for your application.

Justin.