Yes, he talked about this. The key point is that you need to understand how pointwise and depthwise convolutions work and specifically how they are applied in a MobilNet “bottleneck” block. They are special cases of the general convolutions we have been doing. In the pointwise case, they are 1 x 1 convolutions. I gave the link above to the lecture that describes depthwise convolutions. One cool thing about MobilNet that it gives us most of the power of normal convolutions with a lot less parameters to train. That’s why the computation of number of parameters is a big deal here and was covered in the lectures in some detail. You need the depthwise lecture I pointed to above and then the MobilNet Architecture lecture to see how the bottleneck block works. Here’s another recent thread which also discusses how to interpret the slides in that lecture.