What values should I put in my targets whenever they're "don't care" values?

Good day!

I have to train a feedforward net: 6-5-6 I-H-O topology, tansig and logsig activation functions, binary values as targets.

When the value of the SIXTH target is 1, the first five values matter.

However, when the value of the SIXTH target is 0, the first five values are "DON'T CARE" values, or insignificant.

Example: 101111, 100001, 001101, 00000 0, 11111 0, 01110 0

What values should I put in these "don't care" values? Nan? 0's? 1's? Or it wouldn't matter any way? I want to remove the effect of these values in the weight adjustments during training and I'm not sure how to do that in Matlab.

How does the network process inputs with Nan values (I disabled the 'fixunknowns' proccessParam) ?

Thank you!

Answer 1

[Greg Heath]

Make no changes to the net.

Just check outputs for round(y(6,j)) == 0 and act accordingly.

Hope this helps.

Thank you for formally accepting my answer.

Greg

[renz]

I'm really sorry but I still couldn't grasp why it wouldn't matter.

Example: if my DONT CARE values are '11111', i.e. 111110, instead of '00000' or other values, shoudn't those affect the weight adjustments for those values, which are important if it is a success situation?

I want to remove the effects of those first five values whenever it's a fail situation.

Thanks a lot..

Answer 2

[Jason Ross]

You can use the tilde (~) for this.

http://www.mathworks.com/help/matlab/matlab_prog/ignore-function-inputs.html

[Greg Heath]

As I understand it, that ignores certain input matrices.

I think the goal here is to ignore certain columns in those matrices.

I would just search for the 6th zero in the target matrix and remove that I/O pair from training.

Can you give some indication of what the outputs are supposed to represent and why some are to be ignored?

[Jason Ross]

Ahh, yes -- I see the difference. Thanks for the clarification.

[renz]

The sixth binary target indicates a success(1) or fail(0) situation. If a fail, the first five(which determines some value related to a success) are insignificant.

If I remove the I/O pair having the 6th binary target as zero, will the network still be trained to predict a fail situation? From my understanding, it is still necessary. I think I just need to ignore the first five binary targets, not the whole column .

Will the ~ do it?