Alongside the pursuit of accuracy, a second current emerged: making models light enough to run on smartphones and embedded devices. This is where Part I connects directly to what we are doing now.

Depthwise Separable Convolution

An ordinary convolution processes the spatial directions (height and width) and the channel direction (colors, kinds of features) at the same time. With M input channels, N output channels, and a 3×3 kernel, you need 3×3×M×N weights.

Split that into two stages. First convolve each channel independently with a 3×3 kernel (depthwise). Then mix across channels with a 1×1 convolution (pointwise).