True. And it's scaling down properties are much better than any other network I've played with (not an expert). I could run MNIST benchmark on a ESP32 board.

I also liked the Convex Concave trick in the paper. The guarantee that at every step you are closer to the minima is very nice.