Patternnet for multi-dimensional classification

Hello,

I'm trying to develop a neural network for classification of 2 non mutually exclusive outputs, based on 1 input.

Using the standard parameters for the patternnet, I had in the 2 outputs numbers between 0 and 1, summing up to 1. It looked like it was working as they were mutually exclusive, even thought I had in my training matrix cases with both outputs in state 1.

I changed the output transfer function to logsig and removed the default { -1, 1} mapminmax output transformation (after reading this in another topic), but with this I had a very poor training performance...

Any comments would be very appreciated! Thanks.

Answer 1

[Greg Heath]

The training target should have nonnegative entries that sum to 1 and can be interpreted as prior probabilities.

0 and 1 are only used if the classes are mutually exclusive.

The output target should also have nonnegative entries that sum to 1 (e.g., via SOFTMAX) and can be interpreted as posterior probabilities.

Hope this helps.

Thank you for formally accepting my answer

Greg

[Leonardo Campos]

Greg, thank you very much for the quick answer!

Sorry, but I still can't understand how can I workaround with the targets being nonnegative entries that sum to 1, for classes that are not mutually exclusive...

I mean, when I have both of my outputs activated (i.e., target [1 1]), the patternnet with standard parameters gives me something like [~0.5 ~0.5], which is basically the same output as if my target was [0 0]. How can I distinguish these two cases?

[Greg Heath]

ERROR:

TARGETS AND OUTPUTS SHOULD BE NONNEGATIVE UNIT SUM COLUMNS!

1. Input targets with the correct format. Your targets above are NOT CORRECTLY FORMATTED!

2. Use the softmax output function

Hope this helps.

Greg

[Leonardo Campos]

Thanks, Greg. I'll try this out!