depthwise separable convolution

Depthwise separable convolution factorizes standard convolution into depthwise convolution (applying one filter per input channel) followed by pointwise convolution (1×1 convolution for channel mixing), dramatically reducing computational cost and parameters. Standard k×k convolution with C_in input channels and C_out output channels requires k²·C_in·C_out parameters and operations per spatial location. Depthwise separable convolution uses k²·C_in parameters for depthwise (one k×k filter per channel) plus C_in·C_out parameters for pointwise (1×1 convolution), totaling k²·C_in + C_in·C_out parameters—approximately k² times fewer. The factorization separates spatial filtering from channel mixing, which works well empirically despite being a strong architectural constraint. Depthwise separable convolutions are the foundation of efficient architectures like MobileNet, EfficientNet, and Xception, enabling mobile and edge deployment. The approach maintains competitive accuracy while reducing FLOPs by 8-9× for 3×3 kernels. Depthwise separable convolutions represent a key innovation in efficient neural architecture design.