What are neural networks and deep learning?
The Harvard Chan 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 neural network is a mathematical model built from connected layers of calculations. Its internal numerical settings are adjusted as it learns from examples. Deep learning uses neural networks with multiple intermediate layers, allowing information to be transformed repeatedly before the system produces an answer.
The terminology is inspired by biological nervous systems, but the comparison is limited. These computational units are not miniature human brains. Calling a network “neural” does not establish consciousness, human-like understanding or a faithful simulation of biological thought.
Imagine a system recognizing a handwritten digit. Its starting information might be the brightness of each tiny picture element, or pixel. The first layer combines some of those numbers. Later layers combine the results of earlier calculations. The final calculation produces scores for possible digits. With suitable training, the successive transformations can capture distinctions useful for recognizing the number.
Connections between calculations have adjustable values called weights. Other adjustable values can shift when a unit responds strongly. Together these are parameters: the internal numbers learned during training. A layer is a group of calculations at a particular stage; an intermediate or hidden layer sits between the original input and the final output. “Hidden” describes its position, not a deliberate attempt to conceal it.
During training, a learning procedure measures how far the network’s output is from the intended result and changes parameters to reduce that error across examples. The changes can affect many layers. A larger network may offer more capacity to represent complicated patterns, but size also brings greater requirements for computing power, data and careful testing.
Deep learning is used with text, images, sound and many other kinds of information. Different network designs organize their calculations differently. A transformer, for example, is a design that lets pieces of information influence one another through attention calculations. Neural networks are a broader category than transformers.
A neural network could be one candidate for a clinical prediction study on structured electronic health record tables, but complexity is not the starting requirement. Once the decision point and outcome are properly defined, a simpler regression or tree-based model is a vital baseline comparison. Does the deeper network improve performance on genuinely new patients, and can the service act on that improvement? More layers do not repair a misleading outcome label. The brain analogy introduces the idea; it does not confer clinical understanding.
If you would like to find out more about AI and improve your leadership skills, consider 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