Pooling layers are commonly put in between convolutional layers; they are meant to reduce the volume of data being passed around, therefore reducing the number of parameters, as well as reducing the amount of computation required by the network. In this case, we are pooling numbers together by taking the maximum over a given region of numbers.

These layers also work similarly to the convolutional layers; they apply on a predetermined grid and perform the pooling operation. In this case, it is the maximum operation, so it will take the highest value within the grid.

For example, in a max pooling operation on a 2 x 2 grid, the first cell of output will come from the top left, as follows:

And with a stride of two, the second will come from the grid shifted right two rows, as shown here: