Bengio, Hinton and LeCun receive the 2018 ACM Turing Award

A field that had nearly emptied out

A neural network is a computing system built from layers of simple units whose settings are learned from examples rather than written by hand. The idea was introduced in the 1980s, but it spent years on the margins. By the early 2000s, Yoshua Bengio, Geoffrey Hinton and Yann LeCun were among a small group of researchers still committed to the approach ^[1]. The methods that group kept developing are now the dominant paradigm in the field ^[1].

The announcement

ACM announced on 27 March 2019 in New York that the 2018 ACM A.M. Turing Award would go to the three ^[1]. The citation honoured conceptual and engineering breakthroughs that made deep neural networks a critical component of computing ^[1].

The Turing Award is often called the "Nobel Prize of Computing". It carries a $1 million prize, with financial support provided by Google, Inc. ^[1]. The recipients were to be formally presented with the award at ACM's annual awards banquet on Saturday 15 June 2019 in San Francisco ^[1].

At the time of the announcement, ACM described Bengio as Professor at the University of Montreal and Scientific Director at Mila; Hinton as Google VP and Engineering Fellow, Chief Scientific Adviser of the Vector Institute and University Professor Emeritus at Toronto; and LeCun as Professor at NYU and VP and Chief AI Scientist at Facebook ^[1].

What each of them was cited for

ACM's citation named particular pieces of work for each of the three.

For Hinton, it pointed to the 1986 backpropagation paper written with David Rumelhart and Ronald Williams, the 1983 Boltzmann Machines work with Terrence Sejnowski, and the 2012 AlexNet improvements that used rectified linear neurons and dropout ^[1]. Backpropagation is the procedure that lets a network work out how much each of its internal settings contributed to an error and adjust it accordingly; it remains the basic training method behind most of the field.

For Bengio, it pointed to probabilistic models of sequences from the 1990s, the 2000 paper "A Neural Probabilistic Language Model", which introduced high-dimensional word embeddings, the attention mechanism developed by his group, and generative adversarial networks, developed with Ian Goodfellow ^[1]. Word embeddings represent words as vectors of numbers learned from text instead of being defined by hand; attention lets a model weigh which parts of its input matter most at each step of its output; a generative adversarial network sets one network to produce samples and another to judge them.

For LeCun, it pointed to convolutional neural networks developed in the 1980s, an early version of backpropagation, and the modular automatic-differentiation vision proposed with Leon Bottou ^[1]. Convolutional networks reuse a small set of filters across an image instead of connecting every input to every output, which makes them practical for pictures and other grid-like data.

Why it matters

ACM president Cherri M. Pancake said the technologies are used by billions of people, and that advances in natural language processing and computer vision were not possible just 10 years earlier ^[1]. The award thus marked how far a line of research that had once been a minority pursuit had moved: the network architectures and training tricks named in the citation are now ordinary parts of widely used software.

The formal presentation was set for the awards banquet on 15 June 2019 ^[1]. The three shared a single award for a body of work each had built partly with the others in view — a recognition of a research programme as much as of three individual careers.