What neural network architecture should I be using for matching dissimilar types of input data such as a crystal's sampled absorption sprectrum (30 channels) along with a parameter such as its conductivity and to match these to a general database.

I am trying to use a MLP neural network to match the absorption properties of a material, measured over 30 spectral intervals, and one other parameter such as the doping level in that material and to match these patterns to a known database of "good" or "bad" sets of value. Naively, I have just added a 31st input to the input layers with the numerical doping value, however, this does not seem to be too smart. Is there a better neural network architecture such as a fully connected or direct connection between non-adjacent layers that will better suit my problem? Does Matlab have training algorithms for such architectures as presumably the back propagation of error approach would need modification?

Answer 1

[Greg Heath]

Use patternnet for classification and pattern recognition.

Practice first on one or more of the classification nndatasets

help nndatasets

Hope this helps.

Greg