Appendix We first provide additional elements to corroborate our findings: alignment measurement (Section

We report values measured at the deepest DFA layer. Table A.1: Alignment cosine similarity (higher is better, standard deviation in parenthesis) of Table A.2: Alignment cosine similarity (standard deviation in parenthesis) of various graph convolutions architectures as measured on the Cora dataset. We compare DFA to BP, but also to shallow learning-where only the topmost layer is trained. On a simple task like MNIST, a shallow baseline may be as high as 90%. Furthermore, the network is cut down to 3 layers of half the width of NeRF, and no coarse network is used to inform the sampling.