0

I use PyBrain with a simple ANN (64)->(40)->(10) to classify 8*8 digit images. I have managed to train the network up to 96% accuracy, however I noticed that the weights - initially randomized - do not change very much. Is it normal/ an issue ?

Thanks !

Ex : before training 
[-1.56943502  2.15054767 -0.38189152 ..., -0.20049776  1.57006366
  2.00091068]
Ex : after training
[-1.56943502  2.13482924 -0.54452633 ..., -0.23001347  1.59548629
  2.00118807]
Julien
  • 13
  • 5

1 Answers1

0

In and of itself, it's not an issue. If NONE of the weights change much, I'd wonder if I was missing something. However, if some of them don't change much but a few do change some, that just means the ones which didn't change weren't terribly important for this problem set.

The best test, of course, is to test it on additional sample problems from the same population, and see if it makes correct predictions for problems it didn't train on.

rossdavidh
  • 1,966
  • 2
  • 22
  • 33
  • You might also check whether or not a smaller network can just as adequately learn this problem set. Maybe you only need a few neurons for this particular one. – rossdavidh Jun 26 '15 at 13:15