Unable to understand the 0% training for 4th class

The 0% training for 4th class is causing a raise in the error percent. Pleas help me understand why do I get a 0% here. How can I fix it!!

[Greg Heath]

Are your target columns obtained from eye(6)?

What are the row sums and column sums of

 1. target matrix
 2. output matrix

For the count confusion matrix I get

CC =

   130     4     1    20     0     1
    12   134     0    44     0     1
     1     1   135    68     2     1
     0     0     0     0     0     0
     0     1     1     5   142     6
     1     1     2     3     2   138

sumCC =

   144   141   139   140   146   147

sumCCp =

   156   191   208     0   155   147

Therefore, it is not a case of unbalanced priors.

Need to see code.

[Harsimrat Parmar]

I am using nprtool to train my network and for the target I create a file without using eye(). The file I use is created as follows (for 2 classes):

All rows that have input A have a corresponding target of 1,0 and all input rows of type B have a target row of 0,1. I take a transpose when I upload the input and target files.

The network has 899 input features, 30 hidden nodes and 6 output nodes.

[Harsimrat Parmar]

I am still stuck with this problem, looking for a solution! Thank you

Answer 1

[Greg Heath]

1. I do not understand your explanation. Since you have 6 categories, your target columns should come from eye(6).

2. Correct your target structure

 [I N ] = size(input)  % [ 899 857 ]
 [O N ] = size(target) % [ 6   857 ]

3. 889? Whoa ... What is the input data? N =857 examples cannot adequately represent 899 dimensional data.

4. Serious input dimensionality reduction is required. For classification, try PLSregress instead of PCA.

5. Be careful of overfitting the net with too many weights.

The current number of unknown weights is

 Nw = (I+1)*H+(H+1)*O = 900*30+31*6 = 27,186

whereas the available number of equations is only

    Neq = N*O = 5142

and the default number used for training is only

 Ntrneq = N-2*round(0.15*N) = 3,600

5. My advice is

   a. A serious input dimensionality reduction followed by 
   b. A serious attempt to reduce the number of hidden nodes without sacrificing performance.

Hope this helps.

Thank you for formally accepting my answer

Greg