How to conduct a sensitivity analysis in neural network model with Nominal Data

Hello world,

I have built a neural network for pattern recognition (patternnet) and i would like to conduct a sensitivity analysis, however, the main problem is that I have only Nominal Data as input and that makes the noise addition procedure kind of mistaken. Any ideas?

Best regards,

Andreas Kampianakis

Answer 1

[Greg Heath]

Your net is probably so dumb that it doesn't know the inputs are supposed to be nominal (I assume you mean 1s and 0s).

So, plotting performance vs magnitude of added Gaussian noise should be useful.

Hope this helps.

Greg

[Andreas]

How should I build the input in order to be recognized as nominal.? Just like the outputs (1-of-c?

How can a have a specific number of relative importance of each imput through Gaussian noise?

It is always an honor Dr. Heath Best regards,

Andreas Kampianakis

[Greg Heath]

You said you already have nominal inputs. Just add the noise and plot your error vs either noise level or Signal-to-Noise-Ratio.

Greg