how to calculate contribution of each input

omreko 101 Sep 11, 2005 at 11:37

i am using backpropagation with 30 inputs, 2 hidden layers and 4 outputs and i need to know how to calculate the contribution of each input.

3 Replies

Please log in or register to post a reply.

NomadRock 101 Sep 11, 2005 at 18:36

The contribution of each input is the weight you assign to it. These weights are what you change when you train your network.

If you mean what contribution each beginning input makes on the final output then it is a bit more tricky. If this is the case I assume you are asking how much you should change the weights on the lowest level when all you know is what the output of the final level should be.

There are several algorithms for figuring the weight change and they all revolve around the same idea. If your output is too high you lower the weights from the nodes that are high and raise the weights of the nodes that are low. you follow the same logic back to the next level. for those nodes that are too high you not only lower their weight, but you also adjust the weights to them just as if they were an output that was too high. This process can be continued all the way to the origional inputs.

If this also is not what you meant, then im afraid you will have to try to explain again what your trouble is.

omreko 101 Sep 12, 2005 at 03:54

thank you for your answer,
actually , i have developed my ANN model and testing it, i need to know which input has the most effect in the output, also the effect of other input in the output, the result is expected as a percentage of each input, this is called contibution, and is shown in the commercial ANN programs.

awood 101 Sep 13, 2005 at 19:21

I believe what you are looking for is referred to as Example Influence.

Mark White at NCSU has some slides available online, where he covers this. He has some videos of his lectures (scroll down to his lectures from 2003), where you can listen to him give more information on the slides, as well as some entertaining political digressions.

His pdf notes on example influence can be found here: (see slide 33)

Lecture videos can be found here:

Its been a while for me, and I’d have to jog my memory and look at some old code. If you are still having trouble, I can take a deeper look into it.