What is deep learning?
Deep learning is machine learning with neural networks in many layers, able to learn complex patterns in text, images and sound.
The word deep refers to the number of layers in the network. Each layer builds on the one before, from simple features to more complex patterns.
The technique is behind most of today's AI. Language models, image recognition and speech recognition all build on deep learning.
Deep learning needs a lot of data and computing power to train. Using a finished model needs far less.
How it works
A deep network takes in raw data, such as the pixels of an image or the tokens of a text. The first layers respond to simple patterns. Later layers combine them into more complex concepts. No one decides which patterns to look for. The network finds them itself during training.
Training works by letting the network guess, comparing the guess with the right answer and adjusting. This repeats across vast numbers of examples. That is why the largest models are trained by a handful of organisations, while most others use ready-made ones.
Why it matters
Deep learning made it possible for computers to handle free text, speech and images. Most of an organisation's knowledge lives in exactly those forms, in documents, emails and conversations. That is why the technology sits behind most AI tools that are useful in daily work.
What to keep in mind
Deep models can perform very well but are hard to explain. When a decision has to be justified, as in public administration, it is rarely enough that the model was right. You then need explainable AI, or a way of working where a person makes the call with AI as support.
What it means for you
You rarely need to train your own models. Most of the value comes from using finished models on your own documents.
Common questions
How is deep learning different from other machine learning?
Deep learning uses neural networks with many layers and learns patterns directly from text, images and sound. Other machine learning often relies on features someone has chosen in advance.
Why is deep learning hard to explain?
The knowledge is spread across a vast number of weights, so no single rule can be pointed to behind an answer. Read more about explainable AI.
Related terms
Read more
All terms Updated