1

When I trained a backpropagation neural network in MATLAB, I normalized the dataset with the z-score. When I use the trained neural network in Simulink I always normalize the new inputs with the same mean and standard deviation and it gives me good performance, I want to know if what I did is good? enter image description here

MarianD
  • 1,493
  • 2
  • 8
  • 17
  • its better to have normalized data because it may helps to prevent overfiting and helps to model convergence –  Aug 28 '20 at 18:26

1 Answers1

0

So first to clarify backpropangation is not a certain kind of NN rather than just a certain method to derive the gradient for some input. To answer your question, the simplest explanation for normalising data in training is to make the process of training more stable numerically speaking. But that also means that the parameters of your network are manipulated s.t. you want to create the best meodel to given input. In this case your input in training was manipulates, so yes you also need to normalise your new inputs if you want to have (if trained correctly) corret preditcions. Maybe write it down in full mathematical abstract form and it should become more clear.

  • thank you so much for your answer, I used the z-score to normalize the training data, Do I use the same mean and standard deviation of training data with the new data ? –  Aug 28 '20 at 13:44
  • yes you should but note that this is only useful if your new data describes the same process/picture/state as the data you used for training –  Aug 28 '20 at 13:47
  • thank you so much, –  Aug 28 '20 at 14:06