Pooling is an operator which acts on a "window" of data (typically a square in 2D, a cube in 3D, and so on), returning a single representative value for the window. For example: average pooling returns the average; max pooling returns the maximum. Pooling operators are often used in convolutional neural networks to reduce the size of the feature maps -- e.g. 2x2 pooling in 2D reduces each side of a feature map (i.e. downsamples it) by a factor of 2.