In this paper, we study the importance of pruning in Deep Networks (DNs) and motivate it based on the current absence of data aware weight initialization. Current DN initializations, focusing primarily at maintaining first order statistics of the feature maps through depth, force practitioners to overparametrize a model in order to reach high performances. This overparametrization can then be pruned a posteriori, leading to a phenomenon known as "winning tickets". However, the pruning literature still relies on empirical investigations, lacking a theoretical understanding of (1) how pruning affects the decision boundary, (2) how to interpret pruning, (3) how to design principled pruning techniques, and (4) how to theoretically study pruning. To tackle those questions, we propose to employ recent advances in the theoretical analysis of Continuous Piecewise Affine (CPA) DNs. From this viewpoint, we can study the DNs' input space partitioning and detect the early-bird (EB) phenomenon, guide practitioners by identifying when to stop the first training step, provide interpretability into current pruning techniques, and develop a principled pruning criteria towards efficient DN training. Finally, we conduct extensive experiments to show the effectiveness of the proposed spline pruning criteria in terms of both layerwise and global pruning over state-of-the-art pruning methods.

We connect a large class of Generative Deep Networks (GDNs) with spline operators in order to derive their properties, limitations, and new opportunities. By characterizing the latent space partition, dimension and angularity of the generated manifold, we relate the manifold dimension and approximation error to the sample size. The manifold-per-region affine subspace defines a local coordinate basis; we provide necessary and sufficient conditions relating those basis vectors with disentanglement. We also derive the output probability density mapped onto the generated manifold in terms of the latent space density, which enables the computation of key statistics such as its Shannon entropy. This finding also enables the computation of the GDN likelihood, which provides a new mechanism for model comparison as well as providing a quality measure for (generated) samples under the learned distribution. We demonstrate how low entropy and/or multimodal distributions are not naturally modeled by DGNs and are a cause of training instabilities.