difficulty 3/5
Transposed convolution
A convolution run backwards — it scatters the kernel instead of gathering it, turning a small feature map into a larger one. The name is literal, not a metaphor.
Intuition
Convolution and pooling make feature maps smaller. That is what you want for classification, where the answer is one label, and exactly wrong for any task whose answer is the same size as the input.
d2l states the requirement plainly: "in semantic segmentation that classifies at pixel-level, it will be convenient if the spatial dimensions of the input and output are the same. For example, the channel dimension at one output pixel can hold the classification results for the input pixel at the same spatial position" [1, §14.10]. So something has to undo the downsampling, and transposed convolution is the learned way to do it — "also called fractionally-strided convolution ... for reversing downsampling operations by the convolution" [1, §14.10].
The mental model that helps: an ordinary convolution gathers — each output reads a window of the input. A transposed convolution scatters — each input writes a copy of the kernel into a window of the output, and overlapping writes add up. Everything odd about the layer follows from that addition.
Formal definition
How it works
Where it's used
Why & when
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
- Dive into Deep Learning — Transposed Convolutionbook · retrieved 2026-09-14https://d2l.ai/chapter_computer-vision/transposed-conv.html
- U-Net: Convolutional Networks for Biomedical Image Segmentation (2015)paperhttps://ar5iv.labs.arxiv.org/html/1505.04597
- Deconvolution and Checkerboard Artifacts (2016)paperhttps://distill.pub/2016/deconv-checkerboard/