While many researchers celebrated this year’s chemistry and physics prizes, others were disappointed by the focus on computational methods

AI comes to the Nobels: double win sparks debate about scientific fields

While many researchers celebrated this year’s chemistry and physics prizes, others were disappointed by the focus on computational methods

Nobel committees recognized the transformative power of artificial intelligence (AI) in two of this year’s prizes — honouring pioneers of neural networks in the physics prize, and the developers of computational tools to study and design proteins in the chemistry prize. But not all researchers are happy.

Moments after the Royal Swedish Academy of Sciences unveiled the winners of this year’s physics Nobel, social media lit up, with several physicists arguing that the science underlying machine learning, celebrated in the awards to Geoffrey Hinton and John Hopfield, was not actually physics.

“I’m speechless. I like machine learning and artificial neural networks as much as the next person, but hard to see that this is a physics discovery,” Jonathan Pritchard, an astrophysicist at Imperial College London wrote on X. “Guess the Nobel got hit by AI hype.”

The research by Hinton, at the University of Toronto in Canada, and Hopfield at Princeton University in New Jersey, “falls into the field of computer science,” says Sabine Hossenfelder, a physicist at the Munich Center for Mathematical Philosophy in Germany. “The annual Nobel Prize is a rare opportunity for physics — and physicists with it — to step into the spotlight. It's the day when friends and family remember they know a physicist and maybe go and ask him or her what this recent Nobel is all about. But not this year.”

Bringing fields together

Not everyone was troubled, however: many physicists welcomed the news. “Hopfield and Hinton's research was interdisciplinary, bringing together physics, math, computer science and neuroscience,” says Matt Strassler, a theoretical physicist at Harvard University in Cambridge, Massachusetts. “In that sense, it belongs to all of these fields.”

Anil Ananthaswamy, a science writer based in Berkeley, California and author of the book Why Machines Learn, points out that although the research cited by the Nobel committee might not be theoretical physics in the purest sense, it is rooted in techniques and concepts from physics, such as energy. The ‘Boltzmann networks’ invented by Hinton and the Hopfield networks “are both energy-based models”, he says.

The connection with physics became more tenuous in subsequent developments in machine learning, Ananthaswamy adds, particularly in the ‘feed-forward’ techniques that made neural networks easier to train. But physics ideas are making a comeback, and are helping researchers understand why the increasingly complex deep-learning systems do what they do. “We need the way of thinking we have in physics to study machine learning,” says Lenka Zdeborová, who studies the statistical physics of computation at the Swiss Federal Institute of Technology in Lausanne (EPFL).

“I think that the Nobel prize in physics should continue to spread into more regions of physics knowledge,” says Giorgio Parisi, a physicist at the Sapienza University of Rome who shared the 2021 Nobel. “Physics is becoming wider and wider, and it contains many areas of knowledge that did not exist in the past, or were not part of physics.”

Not just AI

Computer science seemed to be completing its Nobel take-over the day after the physics prize announcement, when Demis Hassabis and John Jumper, co-creators of the protein-folding prediction AI tool AlphaFold at Google DeepMind in London, won half of the chemistry Nobel. (The other half was awarded to David Baker at the University of Washington in Seattle for protein-design work that did not employ machine learning).

The prize was a recognition of the disruptive force of AI, but also of the steady accumulation of knowledge in structural and computational biology, says David Jones, a bioinformatician at University College London, who collaborated with DeepMind on the first version of AlphaFold. “I don’t think AlphaFold involves any radical change in the underlying science that wasn’t already in place,” he says. “It’s just how it was put together and conceived in such a seamless way that allowed AlphaFold to reach those heights.”

For example, one key input AlphaFold uses is the sequences of related proteins from different organisms, which can identify amino acid pairs that have tended to co-evolve and therefore might be in close physical proximity in a protein’s 3D structure. Researchers were already using this insight to predict protein structures at the time AlphaFold was developed, and some even began embedding the idea in deep learning neural networks.

“It wasn't just that we went to work and we pressed the AI button, and then we all went home,” Jumper said at a press briefing at DeepMind on 9 October. “It was really an iterative process where we developed, we did research, we tried to find the right kind of combinations between what the community understood about proteins and how do we build those intuitions into our architecture.”

AlphaFold also would not have been possible were it not for the Protein Data Bank, a freely available repository of more than 200,000 protein structures — including some that have contributed to previous Nobels — determined using X-ray crystallography, cryo-electron microscopy and other experimental methods. “Each data point is years of effort from someone,” Jumper said.

Since their inception in 1901, the Nobels have often been about the impact of research on society, and have rewarded practical inventions, not only pure science. In this respect, the 2024 prizes are not outliers, says Ananthaswamy. “Sometimes they are given for very good engineering projects. That includes the prizes for lasers and PCR.”

doi: https://doi.org/10.1038/d41586-024-03310-8

This story originally appeared on: Nature - Author:Davide Castelvecchi