Pooling layer

Down-sampling operation that reduces spatial dimensions of width and height of successive images or feature maps while retaining important information.

This is meant to reduce computation and memory required by the Convolutional Neural Networks, and to increase invariance of features with respect to small translations.

This makes the network more robust to small shifts or distortions in the input image:

• If an object moves slightly in the image, the pooled feature map remains similar.
• The network does not need to relearn features for every small variation in position.

As it has no trainable parameters, it is not a trainable layer.

Types

• Max pooling: Selects the maximum value from the window
• Average pooling: Takes the average of all values
• Sum pooling: Takes the sum of all values in the window