Controlled evaluations of residual connection and normalization designs. A benchmark belongs here if it holds compute or parameters fixed and asks whether added depth turns into useful computation.
Validation loss, domain NLL and layer-level probes (causal, permutation and angular-distance scores) across aspect ratios 9.1 → 76.0.
See figures ↓Have an evaluation protocol for residual designs or depth scaling? Open an issue or a PR and we'll list it here.
Contribute →Measuring How Residual Connections Enable More Computational Depth.
Keyu Wang*, Yangyi Huang*, Jiale Kang, David González-Martínez, Weiyang Liu, Shiwei Liu
*Equal contribution
Click an architecture to highlight it.