Neuro-Lab: Contribution of individual input
Author: alexfong
Creation Date: 12/23/2007 8:52 AM
profile picture

alexfong

#1
Neuro-Lab Build 7

Anyone has a clue how we can measure the contribution of individual "INPUT" of the neural network to the "OUTPUT" in Neuro-Lab.

profile picture

sourkraut

#2
Since NeuroLab doesn't provide multiple outputs, you have to simulate isolating the result of a specific input in two passes.
1. Connect the input to be tested, and plot it as an indicator against the neural net output.
2. Disconnect the input from the NN, and compare the two plots again.

The difference in the two outputs is the specific input's contribution.

If you want a visual comparison, create two separate NN indicators, one for the first pass, one for the second. Then display them both on the graph.
profile picture

aykuts

#3
Hi, does anyone know if it is possible to report, or at least see, the weight matrix of the trained and locked NN indicator? Or, the testing methodology sourkraut defined in post #2 is the only way to (try and) see the individual contribution of each input?

Thanks in advance for your attention.

Aykut

profile picture

Cone

#4
All the details of the network, including the topology and weights, are stored in xml format in the User Data\NeuroLab folder.
profile picture

aykuts

#5
Found it, many thanks.