What is depthwise separable convolution and why is it efficient?

Depthwise separable convolution splits convolution into depthwise and pointwise operations to reduce computation significantly.

Standard convolution mixes spatial and channel-wise processing together, which is computationally expensive. Depthwise separable convolution breaks this into two steps: depthwise convolution (per-channel spatial filtering) and pointwise convolution (1x1 convolution for channel mixing). This reduces computational cost drastically while maintaining performance, making it ideal for mobile architectures.