Hands-On Machine Learning with C++

Pooling operation

A pooling layer receives small, separate fragments of the image and combines each fragment into one value. There are several possible methods of aggregation. The most straightforward one is to take the maximum from a set of pixels. This method is shown schematically in the following diagram:

Let's consider how maximum pooling works. In the preceding diagram, we have a matrix of numbers that's 6 x 6 in size. The pooling window's size equals 3, so we can divide this matrix into the four smaller submatrices of size 3 x 3. Then, we can choose the maximum number from each submatrix and make a smaller matrix of size 2 x 2 from ...