We've discussed the importance of reducing our dimensional space and how we use convolutional layers to achieve this. We use max pooling layers for the same reason—to further reduce dimensionality. Quite intuitively, as the name suggests, with max pooling, we slide a window over our feature map and take the max value for the window. Let's return to the feature map from our diagonal-line example to illustrate, this as follows:

Let's see what happens when we max pool the preceding feature map using a 2 x 2 window. Again, all we're doing here is returning max(values in window):

1. Return max(0,255,255,0), which gets us 255:

2. Step 2:

3. Step 3:

4. Step 4:

By max pooling our feature map with a 2 x 2 window, we've knocked a column and a row off, getting us from a 3 x 3 representation to a 2 x 2—Not bad!

There are other forms of pooling as well—average pooling and min pooling, for example; however, you'll see max pooling used most often.

Next, we'll discuss flattening, a step we'll perform to turn our max-pooled feature map into the right shape for modeling.