Skip to main content

What is a convolutional neural network?

The AI Glossary

In our classes, we often pause to explain the ideas behind AI and the language used to describe it. These posts grew out of those conversations. They’re for anyone who would like to understand AI better, with no technical background assumed.

A convolutional neural network, or CNN, is a neural network that applies learned pattern detectors across an input. It is particularly associated with images, but related designs also process one-dimensional signals and three-dimensional data.

For an image, a small collection of learned numbers is applied to a local region, then reused at other locations. This collection is called a filter or kernel. Its output forms a feature map: a numerical indication of where a learned pattern appears across the image.

Reusing the same filter is called weight sharing. It allows the network to look for a pattern in different locations without learning a completely separate detector for each one. Early layers may capture simple patterns, while later calculations combine information into more task-relevant representations. These descriptions are useful intuitions, not a guarantee that every layer has one easily named meaning.

Many CNNs also reduce the size of their intermediate representations. Pooling is one way to do this, such as summarizing a small region using its largest value. Other designs reduce size through the convolution itself. Pooling is common, not a defining requirement of every CNN.

These choices can make image processing efficient, but they do not make a model automatically insensitive to rotation, scale or distortion. Reusing filters helps with location-related structure; performance under other changes still needs to be learned or tested. Training with altered examples can help, provided those alterations remain appropriate for the task.

In clinical imaging, a CNN may identify a finding, outline a structure or assess image quality. The same architecture can support very different claims. A model trained to classify an image should not be assumed to locate the relevant finding correctly. Changes in scanner settings, patient positioning and the prevalence of findings can affect performance. CNNs are an established and useful design, but choosing an architecture is only one part of defining the data, evaluation and clinical role of a system.

Sources: Goodfellow, Bengio and Courville — Deep Learning, Chapter 9: Convolutional Networks

To learn more, take a look at the AI in Health Care Certificate of Specialization from the Advanced Learning Academy at Harvard T.H. Chan School of Public Health. You can take an individual course or work toward the full certificate.


AI is a fast-moving field. This explanation reflects our class discussions and related material at the time of writing. Terminology, capabilities and tools may change.

For suggested additions or corrections, please email me at tpanch@hsph.harvard.edu.

Dr. Trishan Panch
Program Director, AI in Health Care
Harvard T.H. Chan School of Public Health — Advanced Learning Academy

Lead with AI in Health Care

The AI Pragmatist is a monthly newsletter from ALA faculty, with new explainers & essays from the Advanced Learning Academy