BPNN number of hidden layers and learning curve

[seanshpark]

For a XOR learing system, 2 input 2 output and N(>=1) hidden layers.
How does the learing differ by number of layers?
How do we measure? Will it be valid?

[seanshpark]

With

  alpha_ = 0.90;  // momentum factor
  eta_   = 0.25;  // learning rate
  gain_  = 0.50;  // gain of sigmoid function

and 20 neurons,

1. 2 / 20 / 2
2. 2 / 10 / 10 / 2 : good
3. 2 / 7 / 7 / 6 / 2 : doesn't seem to learn
4. 2 / 5 / 5 / 5 / 5 / 2 : this too doesn't seem to learn

[seanshpark]

Read https://www.cs.toronto.edu/~hinton/absps/JMLRdropout.pdf