The idea in plain language.
Keep a direct path for information and gradients through a stack of residual blocks.
How it works
Residual networks add a learned transformation to a shortcut signal. This paper analyzes how identity mappings let forward signals and backward gradients move directly between blocks. Its experiments motivate a revised residual unit that moves normalization and activation before weight layers. The aim is to preserve the shortcut path while learning useful changes around it.
What to keep in mind
The evidence comes from specific image-classification architectures and benchmarks. A deeper network is not automatically better for every task or budget.
Source: Identity Mappings in Deep Residual Networks. The original manuscript contains the methods, experiments, figures, and references. An arXiv posting date may follow an earlier conference publication. Read the linked record for version history.