midCNN
What are skip connections and how are they different from residual connections?
Updated May 15, 2026
Short answer
Skip connections directly pass activations across layers, while residual connections specifically add inputs to outputs to learn residual mappings.
Deep explanation
Skip connections are a general concept where outputs from earlier layers are passed to later layers. Residual connections are a specific type where the input is added element-wise to the output of a block. This enables learning residual functions and improves gradient flow. ResNet is built on residual connections, while architectures like U-Net use skip connections for feature reuse.
Real-world example
U-Net uses skip connections for medical image segmentation to preserve spatial details.
Common mistakes
- Confusing concatenation-based skip connections with additive residual connections.
Follow-up questions
- Why does U-Net use skip connections?
- What is concatenation vs addition in CNNs?