CNN (Convolutional Neural Network)
A network that slides small filters across an image to find local patterns — the architecture that made computer vision work.
When not to use it
- On non-spatial data. The locality prior is the point; applying it to tabular columns, where neighbouring columns mean nothing, is just an odd dense network.
- At very large data scale, where the prior becomes a ceiling. Given enough images, a vision transformer can learn better structure than you assumed.
- When you need global relationships from the first layer. CNNs build receptive field gradually; some tasks want everything attending to everything immediately.
Reach for something else instead
- Vision transformers at large scale or when you need global context early — and only if you have the data.
- Classical computer vision — thresholding, template matching, edge detection — which is still unbeaten in controlled conditions like a fixed factory line.
- A pretrained model via API if the task is common. Most teams don't need to train anything.
Sources & further reading
- LeCun et al. (1998), Gradient-Based Learning Applied to Document Recognition — LeNet, and the origin of the whole approach.
- Krizhevsky, Sutskever & Hinton (2012), ImageNet Classification with Deep Convolutional Neural Networks — AlexNet, the result that started the deep learning era.
- Liu et al. (2022), A ConvNet for the 2020s — CNNs rebuilt with transformer-era training recipes, and the argument that much of the gap was methodology.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Training from scratch on a few thousand images. Fine-tuning a pretrained model will beat it, in less time, almost always.
- Skipping augmentation, then adding layers to fix the resulting overfitting. Augmentation is the cheaper fix and usually the better one.
- Validating on data that resembles training data more than reality does. The model looks excellent right up until it meets a real camera.
At a glance
FieldDeep Learning
Core ideaslide learned filters over local patches
Key propertiesparameter sharing, locality
Priortranslation equivariance
Best defaultfine-tune a pretrained ResNet
DifficultyIntermediate
Flashcards for this concept
Question
—
Answer
—
1 / 4
Often compared with
CNN vs. vision transformer — a built-in assumption about images vs. learning that structure from far more data.