Now we have in our minds a conceptual machine that will learn the filters that it needs to apply to extract features from an image. But, at the same time, we don't want the machine to overfit on the learning. A filter that is overly specific to the training data is not useful in real life. If a filter learns, for example, that all human faces have two eyes, a nose, and a mouth, and that's all, it wouldn't be able to classify a picture of a person with half their face obscured.

So, in an attempt to teach a ML algorithm to be able to generalize better, we simply give it less information. Max-pooling is one such process, as is dropout (see the next section).

How max pooling works is it partitions the input data into non-overlapping regions, and simply finds the maximum value of that region:

There is, of course, an implicit understanding that this definitely changes the shape of the output. In fact, you will observe that it shrinks the image.