search

seanshpark / help

common help
MIT License
4 stars 0 forks source link

BPNN number of hidden layers and learning curve #6

Open seanshpark opened 7 years ago

seanshpark commented 7 years ago

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 commented 7 years ago

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 commented 7 years ago

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