Hi Younju:
Though I don't have a direct answer to your question, I can share my
experience so far with the Neural Networks. I am using the Neural Networks
toolbox of MATLAB. Basically I am developing a 3 layered NN Architecture
with the implementation of backpropagation algorithm.
It's real easy to code NN programs in MATLAB and it makes the task a lot
easier since MATLAB is excellent in terms of handling matrices.
Anyway, I am happy to see statisticians getting involved in topics such as
Neural Networks. Most of the techniques/algorithms in Neural Networks seem
to be the counter part of popular statistical algorithms.
Here are few things what I have learned from the Neural Networks class which
I am taking this semester:
1) Training data : How much to use ? Should we use noisy data ?
training vs. testing
2) Network Sizing : How many nodes are needed to solve your problem ?
Are there layers always sufficient ?
It's always easy to decide the number of neurons
in the input and output layer. But how many
neurons should we have in the middle layer?
3) Network Pruning : Should we do pruning to get rid of the superflous
neurons ?
4) Learning Rate : This has significant effect on the network
performance. Should it just be 2/max-eigen-value?
5) Momentum : Should we add a momentum term to avoid falling
into any local mimina on the error surface ?
These are some of the key issues that govern the performance of the network.
There are no specific answers to these questions. But you should keep these
in mind and hopefully have a feel for these as you gain expertise.
I would be gald to hear about the experiences of other statisticians as
well.
Regards
Debasmit
---------------------------------------
Debasmit Mohanty
1st Year Graduate Student
Department of Statistics
University of Alabama, USA
http://www.bama.ua.edu/~mohan001/
---------------------------------------
------------ Reply Separator ----------------------
From: Youngju Lee <[log in to unmask]>
Subject: Neural Network
Dear All
I am using nnet program in s-plus but nnet is only one layer. Is there any
body who know how to make 2 or 3 layer neural network using nnet program in
s? or any advice for neural network will be welcome.
Thank you for your consideration
sincerely
Youngju Lee
University of Pittsburgh
Department of Statistics
______________________________________________________
Get Your Private, Free Email at http://www.hotmail.com
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|