What is deep learning?

"Deep" refers to the number of layers in the network, not to deep understanding. A neural network passes data through a stack of layers, and each layer transforms what it receives before handing it on. With many layers, the network can build up complex ideas from simple ones.

Picture an assembly line. The first stations handle raw materials, the middle stations build parts, and the last stations put the finished product together. In an image model, early layers detect edges and patches of colour, middle layers combine them into shapes and textures, and later layers recognise whole objects such as faces or cars.

In classic machine learning, people usually decide which features the model should look at: the size and location of a house, or how often certain words appear in an email. Deep learning skips most of that step. Given raw data and enough examples, the network learns its own features, which is why it works so well on messy inputs such as photos, audio and free text.

The trade-off is cost. Deep learning generally needs much more data and computing power. Training typically runs on GPUs (graphics processing units), chips originally designed for video games that can do huge numbers of simple calculations at the same time.

Neural networks are not new; the ideas go back decades. What changed was that three things arrived together: very large datasets from the internet, affordable GPUs, and better training techniques. In the early 2010s deep networks began to beat older methods at recognising images and speech, and a 2015 review in Nature by Yann LeCun, Yoshua Bengio and Geoffrey Hinton summarised why.

In 2017 researchers introduced the transformer, a deep-learning design that handles long sequences of text especially well. It became the basis of today's large language models, so every chat assistant you use is a deep-learning system.

Deep learning is not a copy of the human brain. Neural networks were loosely inspired by brain cells, but they are mathematical models that work very differently. They are also hard to explain: with millions or billions of numbers involved, it is difficult to say exactly why a network made a particular decision, which is why they are often called "black boxes".

And deep learning is not always the right choice. For small, tidy datasets, such as a spreadsheet of a few thousand customers, simpler machine-learning methods can do as well at a fraction of the cost and are easier to justify to a regulator or a customer.

Face unlock, voice typing, photo search, live captions, translation apps, image generators such as Midjourney and chat assistants such as ChatGPT all run on deep learning. In business tools it powers transcription, document reading, visual product search and most of the AI features added to everyday software.