Neural networks architectural problem

I would like to use MLPs to classify participants (about 600) that will respond to a psychological test. The test has about 300 items, organized in 15 scales.

The decision to use ANNs derives from an hypothesized non linearity of the items scores patterns.

Due to the paucity of the participants I started considering to implement 15 MLPs, instead of a unique MLP for the entire test.

MLP 1~15: each MLP will receive the inputs from the items that belong to a single scale (e.g., MLP "A" will receive 20 inputs from Scale "A" items, MLP "B" from Scale "B" items and so on).

MLP 16: will receive inputs from the scales scores to capture - at least partially - those relations that I have lost by "splitting up" the 300 items.

I will train these 15 + 1 MLPs with the same dataset.

Now, does this procedure make any sense to you or I am addressing the issue the wrong way?

One last question. Could I treat the outputs of these 15 + 1 MLPs as inputs of a final ANN (would be the 17th) that will combine all the results into a single classification decision?

Many thanks to you all.

Answer 1

[Greg Heath]

I have changed three misprints of 16 to 15.

You have 15 MLPs each with 20 inputs

HOWEVER, you have forgotten to state C, the number of classes.

For each of 15 PATTERNNET MLPS

 [ 20 600 ] = size(input1)

 [ C  600 ] = size(target1) %  "C-1" zeros and a single 1 

 [ C  600 ] = size(output1)

For the final merger MLP compare the results of

 1. Rounding the average of the 15 nets
 2. Using a linear classifier

Hope this helps.

Thank you for formally accepting my answer

Greg

[alkaest2002]

Hi Greg. Thank you for your answer. As you didn't criticize it, I do assume that my intention to use multiple MLPS is not considered bad practice. As a matter of fact I did intend 15+1 separate MLPs. 1 for each 15 scales (inputs will be items scores) and 1 that will consider the scales' score (not the items). Then, one last MLP (the 17th) to combine all of these info into a single classification decision. The number of classes should be three for each MLP, I mean the same three for each MLP. I modified my initial request to clarify all of this. Thanks.

[Greg Heath]

I understand. However, as I have stated, your last MLP can probably be replaced by either an average or a linear classifier.