To summarize once again, the pooling layers always leave the third dimension untouched. We usually misunderstand this with a convolution which reduced 1, but that is not happening with max pooling. Usually pooling layers are used to reduce the first two dimensions because it's quite common to use the stride with them, and they usually do a good job of reducing overfitting.

As you will recall from the convolution, we don't specify the nature of the feature to use. We don't tell the neural network to use vertical, horizontal, Sobel, or Scharr filter; instead, we let the neural network figure out what filter is best for the job it's trying to solve. And this means that the values of the filters are just some parameters to be learned by the neural network.

Well, as regards max pooling, there's no parameter to be learned, it's just a transformation of the input matrix to the output matrix. So, one of the questions with max pooling or average pooling is as follows: What's the idea behind them? What are they doing that offers such good results? Well this question is very hard to answer, and use of the pooling layers is mostly based on the experiment results, which show great improvements after using them. But the best thing in my personal opinions is that, they're trying to capture dominant features or dominant activation's from previous layers.

For example, with a nine value, we saw previously that if the nine value will be another part of the image, max pooling will actually preserve that value and use it as an input for the next layer. Somehow, the maximum activation survived on the next layer, and those that are smaller actually kind of cut off. But again, this idea may not work for you and that's OK. It is usually very difficult to explain in literature what they do so well.

Are you excited about using all the building blocks you've learned so far to build a working neural network architecture?