Nonlinear approximation and (deep) ReLU networks (Q2117331)

The authors address the approximation power of ReLU (rectified linear unit) networks and investigate whether such networks really provide a more powerful approximation efficiency than classical methods of approximation. The discussion of this question takes place in the univariate setting as there one has the best chance to obtain definitive results. The authors focus on the advantages of depth, i.e., what advantages are present in deep networks that do not appear in shallow networks.