Glossary
Who is Geoffrey Hinton?
In short
Geoffrey Hinton is a neural network researcher, one of the founders of deep learning and a winner of the 2024 Nobel Prize in Physics.
People · Geoffrey Hinton
MMOOZZGG.com editorsChecked: 1 min read
For a child
How do you learn to ride a bike? You fall, see what went wrong, and next time you turn the handlebars a little differently. For many years Hinton taught computers this way. The computer tries to answer, sees how wrong it was, and corrects itself bit by bit. At first computers did badly, but now programs that recognize cats in photos and understand speech learn this way.
In plain words
Hinton has worked on artificial neural networks for many decades. A neural network is made of many simple nodes, similar to neurons in the brain, and for it to learn something, the connections between them must be adjusted. Hinton and his colleagues described a handy way to do this: the network makes a prediction, calculates the error and passes it back through the layers so that each connection changes a little. In 2012 he and his colleagues trained a large network to recognize objects in photos, and deep learning spread quickly. For this contribution Hinton received the Turing Award and the Nobel Prize in Physics.
For business
Almost everything companies call AI today is trained with methods Hinton developed. Document and product photo recognition, call transcription and the language models in chatbots learn through backpropagation on large sets of examples. For a manager this gives two practical lessons. Model quality depends on the data: if there are few examples or they were labeled with errors, the network will learn the errors too. The cost depends on the amount of training: fine-tuning a ready model on your own examples is usually cheaper than training from scratch, and sometimes a ready model with a good instruction is enough. Which path fits your task is covered in the training for executives.
For a specialist
Hinton’s contribution concerns the training of multilayer networks. In 1983–1985 he applied methods of statistical physics and created the Boltzmann machine, a stochastic network that finds typical elements in data. On 9 October 1986 Nature published the paper by Rumelhart, Hinton and Williams “Learning representations by back-propagating errors”: backpropagation teaches hidden layers to build internal representations. In 2012, at NIPS, Krizhevsky, Sutskever and Hinton presented a convolutional network with five convolutional layers and about 60 million parameters, trained on ImageNet. Hinton shared the 2018 Turing Award, announced on 27 March 2019, with Yann LeCun and Yoshua Bengio. He received the 2024 Nobel Prize in Physics in equal shares with John Hopfield for discoveries that made machine learning with artificial neural networks possible; at the time of the award he worked at the University of Toronto.
Example
A logistics company teaches a neural network to read waybills from photos. Inside is the same method as in the 1986 paper: the model compares its answer with the correct one and adjusts its connections until the error rate is acceptable.
Related terms
Questions and answers
Sources
- Lindau Nobel Laureate Meetings, Nobel Prize in Physics 2024, 08.10.2024lindau-nobel.org
- Nobel Prize, Geoffrey Hinton, Facts, 2024nobelprize.org
- Rumelhart, Hinton, Williams, Learning representations by back-propagating errors, Nature, 1986nature.com
- Krizhevsky, Sutskever, Hinton, ImageNet Classification with Deep Convolutional Neural Networks, NIPS 2012papers.nips.cc
- NYU Tandon School of Engineering, Professor Yann LeCun wins A.M. Turing Award, 27.03.2019engineering.nyu.edu
Where companies use it
Not sure whether your company needs this?
AI solutions map