# Neural Network ART and SOM

1. Apr 6, 2012

### eoghan

Hi!
I attended a class in neural networks and I learnt something about ART and SOM.
I learnt how to implement the learning algorithm, but then? I wasn't told how to use the trained net! I looked in Google and I didn't find much, everybody talks on how to train a neural net, but nobody says how to use it after the learning process.
Do you know a place where I can find examples of (simple) applications of trained ART and SOM?

2. Apr 11, 2012

### Pythagorean

Er, it should be trained on its application, shouldn't it?

3. Apr 11, 2012

### eoghan

Yes, I was given a set of data and I was told how to train the net for those data. Now that I have a trained net for those data, what am I supposed to do? I mean, what can I do with that net?

4. Apr 11, 2012

### chiro

What kind of data are you working with?

Depending on the data and the application, the idea will be to create a neural network that is able to to create some kind of internal representation of a system that should converge to the represent the actual output itself.

If you are trying to create for example a clustering or classification model, then what happens is that the initial data will 'train' the network to create a model for taking an input and classifying either in an unsupervised form (clustering) or in a supervised form (classification) so that with that model, you can feed in some input and the network will classify it according to how it was trained with the training data.

The classification of data (unsupervised or supervised) helps not only for learning but for the identification of patterns. By creating the right classifications (supervised or unsupervised) you are finding patterns and you can use this for learning either in the network, or in another application (like taking the results of the network and using it in another application).

If the neural network is an optimal representation of the model that you are analyzing, then what this network will represent is the minimal representation of the model in it's most compressed form. Most likely the model will not be an errorless 1-1 representation of the model (there will be likely be errors of some sort), but again the point is that the network itself is a good approximation.

How you actually make use of this depends on exactly what you had in mind when you trained the data and created the network and what the actual data is in context.

5. Apr 25, 2012

### eoghan

Uhm... I think I have a deep lack of knowledge of som and art... do you know any good book with applications where I can study them?