SCOP: Scientific Control for Reliable Neural Network Pruning (Supplementary Material)

From top to bottom are input images, features in shallow layers and features in the deep layers. Through standard Schur complement calculation, the semi-definite condition can be derived, i.e., cov[b We train a generative adversarial network [2] to construct knockoff data. The knockoff data are generated by the generator and then sent to the discriminator to verify whether the knockoff condition (Definition 1) holds. The generator and discriminator are optimized alternately and the loss function provided by [4] is adopted. Considering that images are high dimension data, we take multiple pixels as a whole when exchanging elements between real data and knockoff data.