1. For the Bilinear downsampling, things get a bit muddy.

2. As you'll see, almost all CNN Architectures follow the same general design principles of successively applying convolutional layers to the input, periodically downsampling the spatial dimensions while increasing the number of feature maps