can be separable into product of a column and row M2N2→2×N×M2M^2N^2 \rightarrow 2 \times N \times M^2M2N2→2×N×M2 if we can decompose, we can lower the cost of convolution when there is M×MM \times MM×M pixels and N×NN \times NN×N size filter 2D convolution with a separable filter is equivalent to two 1D Convolution