Research · December 10, 2015

Residual connections (ResNet)

An architecture with skip connections was described that made it possible to train much deeper networks.

Why it matters

Residual connections removed a practical limit on depth and are still used in current architectures, transformers included.

The paper addresses image recognition; generalisation to other tasks was shown by later work.

Event date

December 10, 2015

Timeline date

Event date

Verification

Sources gathered automatically · September 16, 2026

Lines

ID

evt-0052

primary Deep Residual Learning for Image Recognition
arXiv · Published December 10, 2015

Related AlexNet at ImageNet
Both concern deep convolutional networks for images; the later one removes a limit on depth.

Related Batch normalisation
Normalisation made training networks as deep as ResNet possible.
Related AlphaFold 2 at CASP14
The architecture rests on deep networks with residual connections and attention.