Computer Atlas

Convolutional Neural Network

Also known as: cnn, convnet

intermediate concept 3 min read · Updated 2026-06-08

A neural-network architecture that uses convolution to detect local patterns — edges, textures, shapes — in grid-like data such as images; the workhorse of computer vision for a decade.

Primary domain
Artificial Intelligence
Sub-category
Computer Vision

In simple terms

A convolutional neural network (CNN) is a neural network designed for images. A plain neural network treats every pixel as an independent input, which throws away the most important fact about an image: nearby pixels are related, and a pattern (an edge, an eye, a wheel) means the same thing wherever it appears. A CNN exploits this by sliding small filters across the image, looking for local patterns and reusing the same pattern-detector everywhere. That makes it dramatically more efficient and effective at seeing than a generic network.

More detail

The defining operation is convolution: a small grid of weights (a kernel, say 3×3) slides across the image, computing a weighted sum at each position to produce a feature map. Key properties make this powerful:

  • Parameter sharing — the same kernel is used across the whole image, so a network learns “an edge detector” once instead of separately for every location. Far fewer parameters than a fully-connected layer.
  • Translation invariance — a feature is recognized wherever it appears.
  • Hierarchy — stacked convolutional layers learn a hierarchy: early layers detect edges and colors, middle layers detect textures and parts, deep layers detect whole objects.

A typical CNN interleaves convolution layers, activation functions (ReLU), and pooling layers (which downsample to summarize a region), ending in fully-connected layers for the final prediction. Landmark architectures — LeNet (1998), AlexNet (2012, which ignited the deep-learning boom by winning ImageNet), VGG, ResNet — progressively went deeper and more accurate.

Since around 2020, vision transformers have matched or beaten CNNs on many large-scale tasks, but CNNs remain ubiquitous for their efficiency, especially with limited data or compute.

Why it matters

CNNs are what made computer vision work well enough to deploy. The 2012 AlexNet result on ImageNet is widely seen as the spark of the modern deep-learning era — it showed that a deep CNN trained on GPUs could crush previous approaches. For a decade they powered nearly all practical image recognition, and the core idea of exploiting local structure with shared filters extends beyond images to audio and time-series data.

Real-world examples

  • Image classification and object detection — photo tagging, medical imaging (spotting tumors in scans), defect detection in manufacturing.
  • Face recognition and the autofocus/scene-detection in phone cameras.
  • Self-driving perception stacks historically built heavily on CNNs to detect lanes, pedestrians, and signs.

Common misconceptions

  • “CNNs are obsolete now that we have transformers.” Vision transformers lead on some large-scale benchmarks, but CNNs are still widely used for their efficiency and strong performance with modest data.
  • “A CNN understands what an object is.” It learns statistical visual patterns; it can be fooled by adversarial pixels or unusual angles that wouldn’t trouble a human.

Learn next

A CNN is a specialized neural network, and it’s the architecture that long powered computer vision.

Read this in a learning path

All paths →

This topic is part of a learning path. Start in context to keep prev/next and progress tracking.

Relationships

Neighborhood

A visual companion to the relationships above. Click any node to visit that topic.