Home › Neural Network Forums › Encog Project › Encog General Discussion ›

Significant input item

ihrytsyuk's picture
Submitted by ihrytsyuk on Thu, 09/02/2010 - 09:44
in

Hello gurus

I'm a newbie to encog and neural networks.

Is it possible to distinguish the significance of input data?
Imagine that we have 3 neurons in input layer, one hidden layer with 6 neurons and 1 neuron in output layer. Is it possible to mark somehow that first entry in input DataSet(I believe it is associated with first neuron in input layer) is less significant for results then second or third one?
I.E. results for {1,1,1} and {2,1,1} inputs are almost equal, but for {1,1,1} and {1,1,2} are significantly different.

We are going to use this ability (if it exists) to decrease number of input DataSet items for network training.

Please suggest, Ivan

‹ Time to give me ideas for Encog 2.5 Error definition in Encoq ›
SeemaSingh's picture
Submitted by SeemaSingh on Sat, 09/04/2010 - 11:30.

I had a need to do something like this once. I experimented with several options, with varying degrees of success.

For one, I always artificially increased that neuron's weights by 10%. I tried different options of always increasing during training, only... runtime only, and both. This worked out okay, at least generally.

What I ended up doing, is actually, rather simple. I just "doubled" the most significant neuron. So if I had three inputs, I added a fourth, which was just a duplicate of the neuron I wanted to emphasis. This ended up being the easiest to implement.

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