Pooling Layers

Input

C x H x W

Hyperparameters

• Kernel size: K
• Stride: S
• Pooling function (max, avg)

Output

C x H’ x W’ where

• H’ = (H - K) / S + 1
• W’ = (W - K) / S + 1

Max Pooling

• Pick Max value in the region
• Often use same kernel size as stride so there is no overlap
  
• Invariance: Max region may still be the same with small shifts
• No learnable parameters
  Similarly, there can be average pooling