Skip to content
CAS

Vision · 2015 · residual learning

ResNet

Residual networks make depth trainable by letting layers learn corrections to the identity: each block computes F(x) + x instead of an unreferenced transform.

ModLensVision

Interactive Diagram

Focus the lens

Click any component to read what it does. Signal direction follows the edges.

skipInputWeighted LayersF(x)AddF(x)+xNext Block

inputInput. Feature map entering the block.

Core idea

If a deeper network should at least match a shallower one, parameterize the difference. Skip connections give gradients a highway and depth stops being the enemy.

Why it exists

Plain deep networks degrade during training long before overfitting; residuals reframe depth as a series of learnable refinements.

Mathematics

y = \mathcal{F}(x, \{W_i\}) + x
Residual block

Data Flow

What moves through the system

  1. 01Input passes through a stem convolution.
  2. 02Residual blocks add learned corrections to their own input.
  3. 03Spatial downsampling stages widen the representation.
  4. 04Global pooling and a head produce the output.

Strengths

  • + Trains reliably at great depth
  • + Skip connections stabilize optimization
  • + The default backbone for a decade of vision systems

Limitations

  • Still local-first: long-range context needs depth
  • Uniform compute across the image

Applications

  • · Backbones for detection/segmentation
  • · Medical imaging
  • · Any task needing a proven visual encoder
Ask CAS