Convolutional neural network (CNN)
A convolutional neural network (CNN) is a neural network architecture that learns spatial features from images using convolutional filters, making it the long-standing workhorse of computer vision for classification, detection, and segmentation. It exploits the local, translation-invariant structure of visual data.
In this article
Building visual or physical AI?
Let's talk.
What is a convolutional neural network?
A CNN is a neural network built for grid-like data such as images. Instead of connecting every pixel to every neuron, it slides small learned filters, or kernels, across the image to detect local patterns, and it stacks these into layers that build up from edges and textures to parts and whole objects. Because the same filter is reused across the whole image, a CNN uses far fewer parameters than a fully connected network and bakes in a useful assumption: nearby pixels are related and an object looks the same wherever it appears. Since AlexNet in 2012 it has been the default architecture for computer vision.
Key takeaways
- A CNN learns spatial features with convolutional filters, purpose-built for images.
- Weight sharing and locality make it efficient and translation-equivariant.
- It was the default vision architecture for a decade and now shares the field with vision transformers.
How it works
- Convolutional layers: extract local features with learned filters.
- Pooling layers: downsample for efficiency and some invariance to position.
- Activation and normalization: add nonlinearity and keep training stable.
- Head: turn features into a classification, detection, or segmentation output.
In practice you rarely build a CNN from scratch. You run a pretrained one and evaluate it.
CNN vs related architectures
| Architecture | Core idea | Strength |
|---|---|---|
| CNN | Local convolutional filters | Efficient on images, strong spatial priors |
| Vision transformer (ViT) | Global self-attention over image patches | Scales with data, fewer built-in assumptions |
| MLP | Fully connected layers | General, but ignores spatial structure |
Why it matters
The CNN made modern computer vision practical, and it is still the right default for most vision problems. Its inductive bias is both its superpower and its ceiling. Locality and translation-equivariance let a CNN learn from modest amounts of data because it assumes structure that images actually have, which is why it wins when data is limited. But those same baked-in assumptions are why vision transformers, which assume less and learn structure from data, overtake CNNs at very large scale. So the CNN-versus-transformer decision is really a question of how much data you have, not which one is newer.
Frequently asked questions
What is a convolutional neural network?
A neural network that learns spatial features from images using convolutional filters, the standard architecture for computer vision.
What is a convolution in a CNN?
A small learned filter slid across the image to detect a local pattern, reused across the whole image.
What is the difference between a CNN and a vision transformer?
A CNN uses local filters and strong spatial priors, a transformer uses global attention and relies more on data than on built-in assumptions.
Related terms
Director of Growth
Jesse Mostipak is the Director of Growth at Voxel51, where the work is helping humans find and trust what the brand knows, and teaching the Google knowledge graph and the LLMs answering on their behalf to do the same. That question, how knowledge gets built inside a system, is one Jesse has been chasing for years. Earlier versions of it ran through a New York City high school science classroom, data science and machine learning, and developer relations at Kaggle, Posit (formerly RStudio), and Baseten. The answer doesn't change much depending on whether the learner is a teenager, a software engineer, or a knowledge graph. Jesse holds a Master's in Education from CUNY Hunter College. LinkedIn
See all articles by Jesse Mostipak
Last updated November 3, 2026
Building visual or physical AI?
Let's talk.
© 2026 Voxel51 All Rights Reserved