What is it?
A CNN looks at an image through small windows, learning filters that detect edges, textures and then shapes. Because the same filter slides across the whole image, it needs far fewer weights and recognises a pattern wherever it appears.
Also known as: CNN, ConvNet, convolution, ResNet, AlexNet, image classification network, convolutional network, pooling layer.
Technical definition
Convolutional layers compute local, weight-shared linear filters followed by non-linearities, often with pooling or strided convolutions and normalisation; residual connections (ResNet, He et al., 2015) enable very deep networks. Inductive biases (locality, translation equivariance) make CNNs data-efficient on images. They are also used for audio spectrograms and 1-D signals. Competing/complementary designs include Vision Transformers (ViT) and hybrids.
Entity type: concept · Freshness class: stable.
When should I use it?
Strong default for image tasks, especially with limited data or compute.
Implementation patterns
- Pre-trained backbone + task head.
- Data augmentation (crop, flip, colour jitter).
- Residual connections.
- Mixed precision on GPUs.
Technologies and frameworks
Commonly associated: PyTorch / torchvision, ResNet, EfficientNet, ConvNeXt. Names are examples of what exists, not recommendations, and projects change quickly.
Practical guide
- Start from a pre-trained backbone.
- Match preprocessing to it.
- Fine-tune with a small learning rate.
- Test on realistic shifted data.
Common failure modes
- Overfitting small image sets.
- Sensitivity to distribution shift (lighting, camera).
- Shortcut learning on background cues.
- Inappropriate input resizing.
Troubleshooting
- Great on validation, poor in the field — Collect field-like data; augment; test on shifted sets.
- Model focuses on background — Rebalance data; use saliency checks.
Contrasted with
- Vision Transformers (ViT) — ViT splits images into patches and uses attention; CNNs use local filters with built-in spatial bias.
- Recurrent Neural Networks and LSTMs — RNNs process sequences step by step; CNNs process grids/windows in parallel.
Questions this page answers
- What is a convolutional neural network?
- How do CNNs work?
- What is a convolution in deep learning?
- Are CNNs obsolete because of transformers?
Typical goals: understand image-model basics; choose between CNN and ViT.
Sources and verification
Status: not independently verified. This entry was authored on 2026-10-06 from general technical knowledge, without live checking of the sources below. Treat every source as a lead to confirm (identifier, title, authorship, current version) before publication or citation.
Freshness class: stable — Core ideas that change slowly (maths, classic algorithms). Re-review yearly.
Canonical sources (unverified leads):
- LeCun et al., Gradient-based learning applied to document recognition, Proc. IEEE (1998) — locator to be identified
- He et al., Deep Residual Learning for Image Recognition (2015) — arXiv:1512.03385
- Krizhevsky, Sutskever & Hinton, ImageNet Classification with Deep CNNs (NeurIPS 2012) — locator to be identified