0

My ANN has two inputs and 16 outputs, sometimes it ignore some classes, even though my dataset has almost the same number of examples by class, as you can see in this image:the points in the 3rd row 4th column is simply ignored, it's very common even if i alter the number of hidden layers, hidden nodes or training examples.

I trained the network a lot of times, and randomly it worked, as in this image:you can see here that each one of the 16 classes is being successfully classified.

In both cases i am using one hidden layer with 32 hidden nodes, 10000 training examples, using stochastig gradient descent and k-fold validation. I couldn't implement regularization, but this problem doesn't look as overfitting.

I just want to know if there is a way to force the neural network to use everyclass to classify, or if there is a problem with my network. I tried googling it but didn't find anything like it. Thanks, if any additional information is needed just ask it.

Romulo Junior
  • 1
  • 1
  • I am still training my english, so sorry if it's not so good – Romulo Junior Nov 20 '19 at 22:38
  • Maybe i could use one vs. all method like in logistic regression? – Romulo Junior Nov 20 '19 at 23:15
  • Hello everybody, first thank you all for all the answers, it really helped me. But i found a way, if anybody is going through this issue. As it works sometimes initializing random weigths, once it work you can use the same network to similar problems, you just need to re-train the working ANN and it will work. As i need to classify 15 situations as in the images, i just trained one that worked to the others. – Romulo Junior Nov 21 '19 at 14:11

0 Answers0