Depth is a scaling axis — if your residuals let it be.
Discover, compare, and contribute to cutting-edge residual connections designed for large-scale deep learning and depth scaling.
Each method comes with a one-line update rule in shared notation, so you can see how the designs differ at a glance.
DepthBench compares designs at a fixed parameter count across width–depth ratios, with open code and checkpoints.
New method, follow-up, or analysis paper? Add one entry to a data file and open a PR, and the site updates.
Every method is one answer to the same question: how should layer l read from, and write to, the representations of the layers before it?
Contribute methods, reproduce benchmark results, or share analysis of how deep layers are really used.