Analyses of Deep Learning (STATS 385)

Pooling layer

Accepts as input:

feature vector of size $W_1 \times H_1 \times D_1$
size of neighbourhood $F$
stride S

Outputs another feature vector of size $W_2 \times H_2 \times D_1$ , where Accepts as input:

$W_2 = \frac{W_1 - F}{S} + 1$
$H_2 = \frac{H_1 - F}{S} + 1$

The pooling resizes independently every channel of the input feature vector by applying a certain function on neighbourhoods of size $F \times F$ , with a stride $S$ .
source
image source

Max pooling

Picks the maximal value from every neighbourhood.

Average pooling

Computes the average of every neighbourhood.

back

Analyses of Deep Learning (STATS 385)

Stanford University, Fall 2019