difficulty 3/5
Receptive field
The region of the input a single unit can possibly depend on — which grows only linearly with depth, and whose influence is concentrated in a Gaussian centre far smaller than the region itself.
Intuition
Pick one number in a deep feature map and ask which input pixels could have changed it. In the first convolutional layer the answer is a 3 × 3 patch. In the second it is 5 × 5, because each of those nine came from a patch of its own. Keep going and the region grows — but only by k − 1 per layer, which is slow.
This matters because a unit cannot recognise something larger than it can see. If the answer depends on an object spanning 200 pixels and the unit's reach is 40, no amount of training fixes it: the information never arrived.
And there is a second, less obvious problem. Not all pixels in that reach matter equally. The ones at the centre are connected to the unit by many paths; the ones at the rim by very few [1, §1]. The region a unit could use and the region it actually uses are different sizes, and the gap widens with depth.
Formal definition
How it works
Where it's used
Worked example
Pitfalls
Check yourself
3 questions
The questions on this page come with access. They are graded on the server, so the answers stay there until you have given yours.
Sources
- Understanding the Effective Receptive Field in Deep Convolutional Neural Networks (2017)paperhttps://arxiv.org/abs/1701.04128
- Dive into Deep Learningbook · retrieved 2026-09-05https://d2l.ai/