This year’s Nobel Prize in Physics has gone to John Hopfield of Princeton University and Geoffrey Hinton of the University of Toronto “for foundational discoveries and inventions that enable machine learning with artificial neural networks.” The pair will receive an equal share of the 11 million Swedish krona prize, worth approximately $1.06 million.
Starting in the 1980s, the scientists drew on concepts in statistical physics to create artificial neural networks, which can store and retrieve information. These networks, inspired by the structure of the human brain, form the basis of many modern applications in artificial intelligence (AI), such as chatbots like ChatGPT and digital assistants like Siri. They are also helping scientists crunch data and look for patterns.
“These artificial neural networks have been used to advance research across physics topics as diverse as particle physics, materials science, and astrophysics,” Ellen Moons, chair of the Nobel Committee for Physics, said at a press conference this morning. “They have also become part of our daily lives, for instance in facial recognition and language translation.”
Neural networks are based on computational layers that can be strengthened or weakened through training, like neurons in the brain. At one end, data are sucked in—a picture, say—and nodes in each layer fire in response to features in the data, such as the color, edges, or lines in a picture. After training on millions of labeled pictures, the layers and nodes are weighted such that the AI can recognize unlabeled pictures.
Hopfield and Hinton pioneered these networks and applied them to early computer vision efforts. Hopfield created an early neural network with nodes that were like atoms in a magnetic material. In trying to recreate input pictures, he optimized the weights of the atomlike nodes by taking a cue from physics and minimizing the overall “energy” of the system. Hinton took up the ideas and built on them to the point they could compete in early computer vision challenges. A key moment came in 2012, when Hinton and colleagues won the ImageNet contest—an annual competition in which object recognition software is challenged to classify thousands of pictures—using a neural net that halved error rates, making incorrect judgments in just 15% of cases.
Speaking on the phone at the press conference, Hinton said he was “flabbergasted” by the news. Modern machine learning tools will lead to big increases in productivity, but he also worries about AI systems run amok. “I am worried that the overall consequence of this might be systems more intelligent than us that eventually take control,” he said.
More: https://www.science.org/content/article/surprise-ai-pioneers-win-physics-nobel
