|
1- In max pooling, do we pick the maximum output of the the pooling region max(activations) ? Or do we take the max(weight * activation + bias) ? 2- What is the choice of transfer function for max pooling network ? If tanh is used, is it a scaled version ? 3- What can be the desired vector shape for tanh, [-1 -1 -1 +1 -1 ....] or [0 0 0 +1 0 ...] ? |
|
Given input, you obtain the next layer's units by doing convolution Wx, then activation(Wx+b) Then you do max pooling by choosing the maximum value of activation(Wx+b) in a neighborhood. I suppose you can also choose the maximum of (Wx+b) before applying the activation function. This saves some computation. thanks for the reply. what do you think about 2 and 3
(Nov 23 '13 at 10:42)
gul
not sure what you mean... max pooling chooses the largest value and deletes the other pixels so the size shrinks.
(Nov 25 '13 at 06:28)
Ng0323
thanks but I do know that 2x2 pooling on 28x28 feature map produce a 14x14 output. Let me explain part 2 of my question. Suppose we want to do max pooling instead of average pooling. Suppose we have two choices for transfer function, tanh and reLU. which one is preferable or it doesn't matter ? If we use tanh, what sorts of complications may arise ?
(Nov 26 '13 at 04:42)
gul
|
