Changing the output of a neural network

Hi there,

I have a problem with the output of a neural network for data classification. The network has 4 target vectors.

Say for example the output (column vector) is

0.1  0  0.4  0.3
0.1  0.52  0.45  0.25
0.25  0.08  0.1  0.1
0.65  0.4  0.05  0.35

and then I round the output to get a one in each column, and get this

0  0  0  0
0  1  0  0
0  0  0  0
1  0  0  0

and then I use vec2ind(output) to return the rows that contain 1

It returns [4 2].

But columns 3 and 4 and are not classified.

How can I make it return the row with the highest probability estimate? So this example would return [4 2 2 4]

Many thanks

Answer 1

[Greg Heath]

>> t = [ 0 0 0 0 % target

       0  1  1  0

       0  0  0  0

      1  0  0  1]

y = [ 0.1 0 0.4 0.3 % output

       0.1   0.52  0.45  0.25

       0.25   0.08  0.1  0.1

       0.65   0.4  0.05  0.35 ]

[ ymax class] = max(y)

t =

     0     0     0     0

     0     1     1     0

     0     0     0     0

     1     0     0     1

y =

    0.1000         0    0.4000    0.3000
    0.1000    0.5200    0.4500    0.2500
    0.2500    0.0800    0.1000    0.1000
    0.6500    0.4000    0.0500    0.3500

ymax =

    0.6500    0.5200    0.4500    0.3500

class =

     4     2     2     4

Hope this helps.

Greg

[Greg Heath]

If posterior probability estimates are desired, use the SOFTMAX activation function. However, if LOGSIG is used a reasonable approximation is to normalize the outputs to have a unity sum:

>> yn = y./repmat(sum(y),4,1)

yn =

0.0909 0 0.4000 0.3000

0.0909 0.5200 0.4500 0.2500

0.2273 0.0800 0.1000 0.1000

0.5909 0.4000 0.0500 0.3500

>> sum(yn)

ans =

1.0000 1.0000 1.0000 1.0000

Hope this helps.

Greg