Pooling (in CNNs)
Convolutional neural networks practically always use pooling.
Pooling is performed over the feature maps resulting from a convolution. The purpose of this is to promote translation invariance, i.e. the feature should be invariant to small displacements in the input image. This also increases the receptive fields of neurons in deeper layers.
You do this by pooling from contiguous regions of a particular size. If the size is too small, you'll not gain barely any translation invariance.
Pooling takes a filter size and a stride argument. But notice that depth is irrelevant because pooling is performed on feature maps i.e. per single channel. So pooling does not affect the amount of channels.
The pooling operation with size $m$ (without stride) reduces feature map size as follows:
$o = floor(i/m)$
Summary:
- Introduces additional translation invariance
- Reduces size of feature maps while keeping highest activations (for max pooling)
- Smaller feature maps leads to a better memory footprint
- Increases receptive field