Li et al. (2017), LIT-014 — ARXIV-1712.09913.
Moved
The claim is at THEORY-011. It is true, it is believed, and it is not
a recommendation: it says why deep networks became trainable, which is what
the THEORY scheme holds. ADR-031 listed this document as one of eight
practices whose titles state a behaviour rather than an instruction and left
them in place pending a decision about what a vacated code should say;
ADR-034 is that decision and this is its first application.
The claim originates in LIT-014. Li et al. compared ResNet-56 with ResNet-56-noshort on CIFAR-scale data under filter-normalised loss-surface plots. As networks get deep, the landscape goes from nearly convex to chaotic, and that transition comes with worse generalisation and, eventually, a loss of trainability. Skip connections prevent it and “promote flat minimizers”, which the paper offers as why they are needed at extreme depth.
There is no practice under it. The instruction the claim would imply — use skip connections — is trunk, and DP-007 keeps the agreed trunk unfiled. That is why this is a move rather than a split.
Do not read Superseded as a judgement on skip connections. On a practice
that status means a better recommendation exists; here it means the document
is in the wrong scheme and the record says where the right one is.