Supplementary Material for Towards a Unified Compositional Model for Visual Pattern Modeling

Assume Hin, Win and Cin denote the height, width and channel number of the input feature maps, respectively; Hout, Wout and Cout denote the height, width and channel number of the output feature maps, respectively; K is the side length of a square kernel. The forward pass of an And-layer, defined in Eq. (1), is of computational complexity O(KCinHoutWoutCout). For the backward pass, the complexities of Eqs. (2)-(4) are the same2: O(CinHoutWoutCout). In comparison, the complexities of the forward and backward passes of a convolution layer are both O(KCinHoutWoutCout). The computational complexity of an Or-layer is the same with that