When I train Multi-Layered Perceptron using Matlab for 6 classes, for one of the classes I get a 0% in GREEN in the confusion matrix. This leads to to a high error percent for the overall training, validation and testing. What is the cause for this and how may I resolve this problem ?

The training data has little more than 200 instances per class. There are a total of 1225 instances.

Target file is created as follows: If there are 2 characters say A and B in the input then for each row in input matrix for character A there is a corresponding row in Target with 0,1 and for every character B there is a 1,0 row in target. So the input nodes in the NN will be 900 and output nodes will be 2 for this example.

Thanksenter image description here

  • $\begingroup$ Can you say more about your situation, your data, & your goals here? I'm not sure this question is answerable in its present form. Eg, what is "MLP"? What is your total N, & how many cases do you have in each class? Etc. $\endgroup$ Commented Apr 4, 2015 at 2:57
  • $\begingroup$ I have now uploaded the image of my confusion matrix that I obtain after training my MultiLayered Perceptron (MLP) Neural Network ! $\endgroup$ Commented Apr 4, 2015 at 21:46
  • $\begingroup$ I am trying to train the Neural Network for handwriting recognition. Currently for only 6 characters. $\endgroup$ Commented Apr 4, 2015 at 21:48
  • $\begingroup$ Thanks for clarifying. What is your total N, & how many cases do you have in each class? $\endgroup$ Commented Apr 4, 2015 at 22:24
  • $\begingroup$ There are approximately 200 instances for each class. There are 900 input nodes and 6 output nodes. I have tried to train with a single hidden layer of 10 nodes, 20, 30, 100, 200, 600 nodes. But the error varies only slightly. $\endgroup$ Commented Apr 4, 2015 at 23:00

1 Answer 1


I tried to reduce the dimensions of the input layer by using a different feature vector. This helped. When the input layer has a large number of neurons then the Neural Network requires a larger number of instances to train and find the best combinations of weights. When I reduce the number of neurons in the input layer then the number the weights that needs to be updated decreases too thus the NN converges.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.