The residual idea
A residual block can write its output as y = x + F(x). The branch F learns a change relative to the input. If F outputs zero and dimensions match, this block returns its input.
A tiny block
In the example F(x) = ax. The output is (1 + a)x. You can change a and see the direct path and residual branch combine. A real ResNet contains many layers and uses more than this scalar operation. The two paths meet at addition: the input can travel directly while the branch contributes a change.
INTERACTIVE EXPLORATION
Compare the paths
For x = 2 and F(x) = ax, adjust the residual parameter a.
Why does this help?
Residual connections helped researchers train deeper image recognition networks. They are not a promise that all deeper networks will train well, and the skip path may require a projection when dimensions change.
Check your understanding
If the residual branch is zero and dimensions match, what is the block output?
It equals the input.