Material Insights · Understanding the Material World

Deep learning

Many-layered neural networks learn to recognise images, speech and patterns from data.

Computing and AI · AI and data · No. 290

  • Digital age
  • Readiness 9 of 9
  • Process
Computing and AI · Technology

An idea only a small group of researchers still backed in the early 2000s now recognises images, transcribes speech and translates text.

Chart: Illustration: Deep learning

Where it comes from

Chart: Map of where it comes from

Rosenblatt showed his perceptron in 1958. Hinton and colleagues made multilayer networks learnable with backpropagation in 1986; by 1998 such networks read bank cheques commercially, and in 2012 Toronto students almost halved the ImageNet error rate.

Three things to know
  1. Backpropagation, the method that trains layered networks, was set out in a 1986 paper by Rumelhart, Hinton and Williams.
  2. In 2012 a network trained on two graphics cards almost halved the error rate in the ImageNet image-recognition contest.
  3. Hopfield and Hinton shared the 2024 Nobel Prize in Physics for work that made machine learning with neural networks possible.
The journey
  1. 1 DataMillions of labelled examples, such as photos tagged with what they show
  2. 2 NetworkMany layers of simple artificial neurons joined by adjustable weights
  3. 3 TrainingBackpropagation nudges every weight to reduce the errors, on GPUs
  4. 4 ModelThe trained weights now encode patterns found in the data
  5. 5 UseThe model labels new images, transcribes speech or flags fraud
Go deeper pick a tab to open it in full