Can target vector of neural network have sum of columns >1?

i have 56 classes ( 1 , 2, .....56) as output. i want to define target vector for neural network. I have converted classes into binary and output neurons are now 6. So for each column the value is >1. Is it ok? Can u help me to write a code to convert 6 output into original classes?

Answer 1

[Greg Heath]

No! Target columns should be columns of eye(56);

 truclassindices = vec2ind(target);
 target          = ind2vec(truclassindices);
 ...
 output          = net(input);
 estclassindices = vec2ind(output);

Hope this helps

Thank you for formally accepting my answer

Greg

[Greg Heath]

CORRECTION:

The above answer ASSUMED that the classes were mutually exclusive.

If the classes are NOT mutually exclusive e.g., tall, dark, handsome (excuse me for blushing), then the nonzero entries are fractions that sum to 1.

By the way, there are two forms of crossentropy: One for mutually exclusive classes and another for non-mutually exclusive classes.

I have advised MATLAB of this but they have not informed me of any upgrades.

I probably have posted something about this.

Search the NEWSGROUP and/or ANSWERS and/or comp.ai.neural-nets using

 greg crossentropy 
 greg cross-entropy 
 greg cross entropy

Hope this helps.

Greg

[Greg Heath]

CORRECTION OF THE CORRECTION

If the classes are NOT mutually exclusive e.g., tall, dark, handsome (excuse me for blushing), then the nonzero entries are fractions

THAT DO NOT HAVE TO SUM TO ONE!