Home › Neural Network Forums › Neural Networks › General Neural Networks ›

Backpropagation

aliaz's picture
Submitted by aliaz on Wed, 03/10/2010 - 07:01
in

Hi,
I trained a FeedforwardNetwork with backparopagation and the error doesn't decrease from the beginning (maybe that's not a probleme. There are 486 input neurons !). But then the error becomes zero without being small enough. And the end of the training, I tested again the input and the results were exactly equal to their Ideal outputs. Do you think that my network has been trained very well, or that there is a problem (because even for other inputs it gives the results +1 or -1)
Thank you for answering.

‹ RNN for DTM project Feedforward networks ›
SeemaSingh's picture
Submitted by SeemaSingh on Thu, 03/11/2010 - 06:44.

But, that is a very large network, do you have the same number of inputs and outputs.

For the error to become "zero", this would mean that the output is exactly matching the ideal output. Which is generally a good thing, unless your sample data is badly skewed.

aliaz's picture
Submitted by aliaz on Thu, 03/11/2010 - 14:32.

As proposed in the book : Introduction to neural networks (Java), I tried to have 2/3 of the number of input plus output layers (near 300 neurons). The problem is that even for test examples, the networks gives the values +1 or -1 for each input which is quit odd because the network is used for face detection. I am not sure that i would be that precise !

aliaz's picture
Submitted by aliaz on Mon, 03/22/2010 - 13:22.

Hi,
I modified my network in order to have less inputs neurons an connexions. Now the error is decreasing (slowlly). I just wanted to now if the momentum and the learning rate depend only on the netwok design (number of nerons etc) or also on the size of the training data ?

Thank you !

Copyright 2005 - 2012 by Heaton Research, Inc.. Heaton Research™ and Encog™ are trademarks of Heaton Research. Click here for copyright, license and trademark information.