What is deep learning?

In plain English

Deep learning is a style of machine learning built on neural networks that have many layers stacked on top of each other — that stacking is what “deep” refers to. Each layer transforms the information a little and passes it on, so early layers pick up simple features and later layers combine them into more abstract ones.

The practical breakthrough is that you no longer have to tell the system what to look for. Older approaches required experts to hand-design the features a model should measure. A deep network works that out for itself from the raw pixels, audio samples or text.

Why it matters

Almost every impressive AI capability of the past decade — speech recognition that actually works, photo-realistic image generation, fluent machine translation, chatbots — comes from deep learning. It scales unusually well: feeding bigger networks more data and more computing power has kept producing better results, which is why AI has advanced so quickly and why training the largest systems is so expensive.

An example

When you dictate a message to your phone, a deep network turns the sound wave into text. Its lower layers detect raw acoustic patterns, middle layers assemble those into speech sounds, and higher layers work out which words and phrases those sounds most plausibly form, given the context of the sentence.

What to watch out for

Deep learning is powerful but hungry and opaque. It typically needs enormous amounts of data and specialised chips, which concentrates capability in well-funded organisations. And because knowledge is spread across millions or billions of numerical settings, nobody can read off exactly why a given answer came out — making errors hard to diagnose and fix. “Deep” also says nothing about depth of understanding; it is purely a description of the network’s architecture.