Studying NN in breadth

Studying NN in breadth

Post by benjamin.c » Sun, 21 Aug 2005 04:56:35


I'm studying for an exam and want to gather as much knowledge about NN
as possible. I'm reading different textbooks, but now I want to
implement different architectures to learn more about the subject.

I have different questions where help will be much appreciatied:

1. Is it good to implement NN on your own and/or to use existing
packages like JOONE?

2. Which problems should one tackle, especially at the beginning? XOR,
Game Nim, TicTacToe, etc.?

3. What Topologies/Algorithms should one focus on? Adaline, Perceptron,
Backprop, Kohonen, etc.?

Thanks in advance,

Studying NN in breadth

Post by Greg Heat » Sun, 21 Aug 2005 14:21:04

That sounds very inefficient and ineffective.
Why don't you concentrate on what will be on the exam?

Existing first, then your own.

I am not familiar with JOONE. I use MATLAB.

Whatever interests you.

Backgroung: PCA, Clustering(unsupervised/supervised, k-means/leader
clustering), Sammon display
Topologies: MLP, RBF & EBF
Algorithms: Rprop, Quickprop, L-M, SCG
Overtraining Mitigation:
Weight decay, Early stopping, Jittering, Bayesian Regularization

Others will have a different opinion.

Hope this helps.



Studying NN in breadth

Post by benjamin.c » Sun, 21 Aug 2005 18:41:45


The problem is I find it hard to find small problems i can solve. If
the problems are to hard I don't want to work on them, because I want
to see results in a short period of time.

What is MLP, EBF, PCA, SCG?

Studying NN in breadth

Post by Greg Heat » Mon, 22 Aug 2005 01:29:20

Use the search engine at

also search in

Hope this helps


Studying NN in breadth

Post by darbeda » Tue, 23 Aug 2005 18:25:49

Dear Benjamin,

I don't know about JOONE, but I have been using Matlab for the last 17
years. Basically if you have some C language skills / familiarity you
can easily work with Matlab and the Neural Networks toolbox in there.
Write a few simple scripts / functions to try the various training
algorithms and network configurations.

You can experiment with feedword networks as well as recurrent networks
for example. The toolbox help is very usefull as it serves as a
tutorial into neural networks.

There are code smaples within the text which you can click on to
excecute and see the results for yourself.

Of course you should try writing your own stuff, but you can use the
existing matlab code there as a basis.

What you get out of the neural net toolbox ultimateley depends on what
you are after. If you want to get your hands dirty with and experiment
with the actual algorithms and traioing functions ( modifying them for
example etc ) you can do that too but you should know what you are

If you want to simply try various data sets, train various network
configurations and training parameters, and observe the results you can
do that very effectively.

As for sample poroblems, look in the toolbox tutorial to see how smaple
problems are presented to the net and how the net is trained, then
practice by writing your own code and running it.


Reza Mostafid