Hands-On Machine Learning for Algorithmic Trading

Max pooling

A common pooling operation is max pooling that downsamples a feature map by using only the maximum activation value from (typically) non-overlapping sub-regions. For a small 4 x 4 feature map, 2 x 2 max pooling would compute the max for each of the four non-overlapping 2 x 2 areas, and output the result.

Alternative pooling operators use the average or the median instead of the maximum, but the latter is most common. Pooling does not add or learn new parameters, but the size of the input window, and possibly the stride, are additional hyperparameters.