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

Feedforward networks

aliaz's picture
Submitted by aliaz on Sat, 03/06/2010 - 12:03
in

Hi,
I m trying to use feed forward networks for face detection. I want to know if it is possible to manage connexions between neurons in your feed forward network. Because when you create a new layer all the neurons are connected.

Thank you

‹ Backpropagation Kohonen Map Question- Preference for Corners of the Map ›
SeemaSingh's picture
Submitted by SeemaSingh on Sat, 03/06/2010 - 13:53.

Is to set individual weights to zero for the neurons that you do not wish to be connected.

The latest version of Encog (2.4) also supports a type of neural network called NEAT, where neurons are connected individually, and there are essentially no layers.

aliaz's picture
Submitted by aliaz on Sun, 03/07/2010 - 05:50.

Hi
I tried to use Encog but I can't find which methode to use to manage connections. The synapse class provides only types that conenct all neurons or weightless sypanses !! So how can I set weights to zero ?

Thank you very much

John Merk's picture
Submitted by John Merk on Sun, 03/07/2010 - 11:30.

My guess is you would have to edit the network file "by hand' to accomplish this. Someone please let me know if I am wrong.

jeffheaton's picture
Submitted by jeffheaton on Sun, 03/07/2010 - 19:51.

You are going to need to get at the weight matrix, which is in the synapse. This is really similar discussion to the one in the other forum about directly connecting neurons. Seems to be a popular feature and one that could defiantly use some helper functions plus enhancements to Encog training. This is defiantly on the list for 2.4.

Just out of curiosity. I really have never created a neural network setup like this. How do you determine which neurons to not connect? Is it trial and error, or is there some method for this.

Jeff

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