Pooling Layer

Pooling layers reduce the size of the input image to reduce the amount of computation and parameters in the network. Pooling layers are inserted periodically between convolutional layers to control overfitting. The most common variant of pooling is 2 x 2 max pooling with a stride of 2. This variant performs down-sampling of the input to keep only the maximum value of the four pixels in the output. The depth dimension remains unchanged.

Figure 6.8: Max pooling operation

in the past, we used to perform average pooling as well, but max pooling is used more often nowadays because it has proven to work better in practice. Many data scientists do not like using pooling layers, simply due to the information loss that accompanies the pooling operation. There has been some research on this topic, and it has been found that simple architectures without pooling layers outperform state-of-the-art models at times. To reduce the size of the input, it is suggested to use larger...