Deep Residual Learning for Image Recognition

TL;DR

Introduces residual networks (ResNets), where layers learn a correction to their input via skip connections, making it practical to train networks hundreds of layers deep.

Why it matters

Before ResNets, very deep networks got harder to train, not better. The residual connection fixed that and became a standard building block, you will find the same skip-connection idea inside Transformers.

Key ideas

• A block learns a residual F(x) and outputs F(x) + x instead of a fresh mapping.
• Skip connections give gradients a short path back, easing the degradation of very deep nets.
• Enabled 100+ layer networks that kept improving with depth.
• The residual pattern generalizes far beyond vision.

Related concepts