What is Convolutional neural network?

What a CNN is

A convolutional neural network, or CNN, is a kind of neural network designed for data laid out in a grid, most often images. It was the dominant approach to computer vision for years before transformer-based models also became common there, and it is still widely used because it is small, fast and effective.

How it works

A photo is a grid of pixel values. A CNN does not look at all pixels at once. Instead it uses small filters, tiny grids of learned numbers, and slides each one across the image. At every position the filter measures how strongly that patch matches a pattern.

Early layers pick up simple patterns such as edges and corners. Later layers combine those into textures, then parts, then whole objects. Between layers, a pooling step shrinks the picture so the network focuses on what matters and not exactly where it sits. The final layers turn all of this into an answer, such as "this looks like a cat" or "this lesion looks like a melanoma."

The filters are not designed by hand. Like any neural network, a CNN learns them during training by adjusting its parameters to reduce its mistakes on labeled examples.

Why the design helps

Sliding the same filter everywhere means the network recognizes a pattern wherever it appears, and it needs far fewer parameters than a network that treats every pixel separately. That makes CNNs efficient to train and cheap to run, even on modest hardware.

Where you see them

• Medical imaging, such as scans and skin photos
• Face unlock and photo tagging
• Quality inspection in factories
• Reading handwriting and documents
• Self-driving perception

A well-known family is Inception, introduced by Google researchers in 2015. A common practice is to take a CNN already trained on a large image collection and apply fine-tuning on a smaller, specialized set, such as images from one medical field.

CNN versus a general chatbot

A narrow CNN trained on one kind of image can beat a large general model at that one job, because every bit of its training went to that task. A multimodal model handles many kinds of input but is not specialized. When accuracy on a single image task matters, test a specialist first, using your own evals.

Related terms

CNNs are one architecture among several. Image generators often use a diffusion model, and many modern vision systems use transformers. All learn from training data and run as inference once trained.