An important part of processing multimedia data for the purpose of machine learning is downsampling it. This provides two benefits:

• It reduces the overall dimensionality of the problem, decreasing the time needed to train the model in a major way.
• Through aggregation, we abstract the unnecessary details in the multimedia data, making it more generic and more representative of similar problems.

Downsampling is performed as follows:

Note that we have replaced every block of four pixels with one pixel, choosing the highest value of the four pixels to be the value of that one pixel. This means that we have downsampled by a factor of four. As we have chosen the maximum value in each block, this process is called max pooling. We could have chosen the average value; in that case, it would be average pooling.