Back to timeline

Deep Residual Learning for Image Recognition

He and colleagues introduce residual connections, enabling very deep networks and influencing modern architecture design.

Architecture

What Happened

In late 2015, the ResNet paper proposed residual (skip) connections, making it possible to train substantially deeper networks by easing optimization.

Why It Matters

Residual connections became a broadly used design pattern across deep learning, influencing architectures in vision and beyond, and contributing to stable training of deep stacks.

Technical Details

Residual blocks add an identity shortcut around transformations, allowing gradients to propagate more reliably and reducing degradation as depth increases.