Residual block in SANSA. An input vector x is first passed through a convolutional layer and a copy of the output tensor is made which consecutively goes through a pair of convolutional layers introducing nonlinearity, all the while preserving the shape of the output tensor. The outcome is then algebraically added to the earlier copy (i.e., a parallel, identity function) and the sum is passed through a nonlinear activation to obtain the final outcome of the block. The latter two convolutional layers thus learn a residual nonlinear mapping. (Note that a zero-padding is applied during all convolutions in order to preserve the feature shape in the subsequent layers through the network.)