Deep Learning Essentials

Pooling/subsampling

The pooling layer progressively reduces the spatial size of the representation to reduce the number of parameters and computation in the network. For color images, pooling is done independently on each color channel. The most common form of pooling layer generally applied is max pooling. There are also other types of pooling units, such as average pooling or L2-norm pooling. You may find some early networks are using average pooling. As Max Pooling typically shows better performance in practice, average pooling has recently fallen out of favor. It should be noted that there are only two variations of max pooling commonly seen in practice: 3 x 3 with stride = 2 (also called overlapping pooling), and even more commonly, ...