How to exactly predict the percentage of individual classes detected in a neural network regression problem

Hi,

I've already trained several neural networks with 5 classes (0-4) as a regression problem. Stopping criterion used was MSE. I have got good results with MSE, R-vale close to 0.99.. etc. and regression plot for all kinds.

Is it possible to show my classes predicted in this regression fitting problem. The problem here is my target vector is (1xN) for different subjects and when I run plot confusion it gives error as the target contains classes as 0,1,2,3,4 instead of only 0 and 1.

To solve this I had created another target vector with all the classes as required and included "confusion" command for training such that I get confusion matrix (5x5)and calculate the results accordingly. But it was not generalizing after so many times I trained.

The regression worked fine but I am not able to cite individual class detection results though on regression plot they are shown exactly fitting on the respective classes. Citing only MSE and R-value doesn't differentiate individual class detection results. Is it ok if I just use MSE and R-vale?

Any response in this regard will be highly appreciated.

Thanks in Advance

[LRBLFAST]

No reply as yet,

Answer 1

[Greg Heath]

You don't have a regression problem. You have a 5-class classification problem.

The target matrix size should be [ 5 N ] consisting of column vectors from the 5-dimensional unit matrix eye(5). Use vec2ind and ind2vec to switch between unit vector and class index representations.

classind = vec2ind(net(input))

Nerr1 = sum(classind ~= 1)

etc.

Greg