ResNet
A very deep CNN that uses shortcut connections, letting each block learn only a correction to its input. It made networks with 100+ layers trainable.
Also residual network · ResNet-50
He, Zhang, Ren and Sun, Microsoft Research, 2015
Instead of every floor of a tall building redrawing the whole plan, each floor just adds small notes to the plan it received. Even with a hundred floors, the original plan stays readable.
Where it fits
+All relationships as a list
Is a kind of / part of
Siblings
Uses / built on
Used by
Evolved from
How it works
By 2015, researchers had noticed something odd: stacking more layers onto a CNN eventually made it worse, even on training data. The extra layers were not overfitting; the network simply failed to train.
ResNet’s fix is the residual connection. Each block takes its input , computes some change , and outputs . If a block has nothing useful to add, it only needs to learn , which is easy. And during backpropagation, gradients flow straight through the ”+ x” shortcut, so they do not vanish across many layers.
With this trick, ResNets with 50, 101 and even 152 layers trained well and won the 2015 ImageNet competition. Residual connections then spread far beyond vision: every Transformer block uses them.
Everyday use cases
- Case 01
Standard vision backbone
Pretrained ResNets are a common starting point for medical, industrial and satellite image models.
- Case 02
Transfer learning
Replace the last layer and fine-tune on a few thousand of your own images.
Strengths, weaknesses and when to use it
Strengths
- Trains reliably at great depth.
- Simple, well understood, widely available pretrained.
Weaknesses
- Matched or beaten by newer architectures at large scale.
When to use itAs a dependable, well-supported image backbone, especially with limited compute.
Compare with
Check yourself
- Q01
What does a residual block compute?
Show answer
B: y = x + F(x)