Artificial Intelligence Didn’t Deserve the Physics Nobel Prize
The science Nobel prizes come in two broad categories: most are given to people who discovered significant true things about the world, while a minority are given to people who devised really good tools.
This year’s Nobel Prize in Physics was awarded to Geoff Hinton and John Hopfield for their work on artificial intelligence. This was a mistake. It’s like declaring the New York Yankees Super Bowl champions, or giving Katie Ledecky a gold medal in the 100-meter dash because she’s a fast swimmer. Giving Hinton and Hopfield the prize sends a message that artificial intelligence is the most exciting thing going on in physics today. It’s not. The neural networks they developed are not themselves a physical discovery, nor have they, as a tool, yet enabled any Nobel-prize-worthy physics.
On the other hand, half of the Chemistry Nobel went to scientists at Google who applied AI to predict how proteins fold. AlphaFold, the algorithm they developed, can, in an instant, solve problems that used to take weeks, months or even years. AlphaFold debuted in December 2020, and revolutionized protein science. It’s a Nobel-worthy tool, but understanding why it’s so effective reveals just how limited artificial intelligence’s ability to make scientific discoveries is.
Plenty of scientists are dismissive of prizes. The true value of their work, they say, is the work itself. The joy of learning something new about the world is its own reward. But prizes matter. They serve as signposts for the public about what scientists consider important, and they inspire researchers to strive for excellence. At their best, they not only commemorate discoveries that have taken place in the past, but inspire and provoke future work.
Physics is about trying to understand why the natural world is the way it is. There is a spiritual significance to comprehending that time and space are relative rather than absolute, as Einstein discovered. Wrestling with the contradictions and counterintuitive consequences of quantum mechanics has changed the texture of modern life and had profound practical consequences. Physics can be celebrated both for its abstract implications and for the concrete changes it brings about in the world, including the discovery of electricity and the ability to manipulate electrical signals on integrated circuits that underpin modern computers.
Every few decades, a new discipline emerges in academia. In the 1940s and 50s, computer science was born of a marriage between math and electrical engineering. Physics was, at best, a guest at the wedding. Computer science — the study of what algorithms can accomplish — is an exciting and worthy subject in its own right. It didn’t exist when Alfred Nobel died, so there isn’t a Nobel. But it has its own prizes, most notably the Turing award, commonly called the “Nobel Prize of Computer Science” which Hinton already won, in 2018.
Disciplinary boundaries exist in prize-giving (and in universities) because it’s impossible to reasonably compare the importance of discoveries in different areas. It’s tough enough to make those value judgements within a discipline. But it simply doesn’t make any sense to say that DNA is more important than the Big Bang, or vice versa. Of course the boundaries between scientific disciplines are not sharply defined. But this year’s prize is nowhere close to any interdisciplinary borders.
To their credit, both physics winners have expressed surprise at being given an award in a field they hadn’t been working in. There is, of course, precedent for this decision: the 2016 literature prize, which went to Bob Dylan. Just like passing over all the novelists, poets and playwrights in the world in favor of an (admittedly terrific) musician and songwriter was a mistake, so too is telling all of the astrophysicists, biophysicists, geophysicists, condensed matter physicists (to name just some sub-disciplines) that their work doesn’t matter as much as somebody else’s achievement in a distinct field, even if there are some loose connections to physics.
An incomplete list of overlooked discoveries: The Aharonov-Bohm effect, in which electric and magnetic fields exert influence on particles even when they don’t exist is tremendously surprising, and worth celebrating. So are metamaterials that bend light in bizarre ways. Quantum computers are remarkable and important (and, unlike neural networks, truly dwell on the boundary between computer science and physics.) Without getting into the question of what physics discovery should have gotten the Nobel nod this year, it’s worth making the case that the prize for physics should be given for work in physics.
Many people think about physics only rarely, if at all. Nobel-worthy physics discoveries might have practical importance, like the x-ray (the very first physics Nobel, in 1901) or laser (1964 physics Nobel) or they might not (gravitational waves, 2017, or black holes, 2020). The money that’s attached to the prizes is certainly nice for those who win, but mostly it serves to get everybody’s attention, to shine a spotlight on something esoteric and to say: this matters.
Artificial intelligence research is obviously consequential. But it has not (yet, at least) contributed any explanations that address the question of why things are they way they are. Neural networks like those pioneered by Hinton and Hopfield are powerful tools for making statistical inferences from large datasets. But they aren’t yet any more than that.
The half of the chemistry prize, given to Dennis Hassabis and John Jumper, who elaborated Hinton and Hopfield’s ideas to create a so-called deep neural network that can correctly predict how proteins fold, is merited. But one crucial reason that AlphaFold is so effective is because there already existed a huge dataset of folded proteins painstakingly decoded by people over the course of the last few decades. AlphaFold could train on this dataset, which has a very clean structure. Sequences of amino acids — the building blocks of proteins — are the inputs, and the outputs are the three-dimensional structure of proteins.
Deep learning algorithms, with their powerful capacity for statistical inference, are likely to make similarly spectacular breakthroughs when there is lots of well-organized data lying around, with a clear relationship between inputs and outputs. But science is only rarely so clean-cut. And though AlphaFold can say how proteins fold, it cannot say why.
To the extent that science is about understanding the reasons behind natural phenomena, neural networks have not yet — and might not ever — make a scientific contribution. Though they are useful for sifting through data, they simply don’t have anything to say about the reasons reality is the way it is. In the citation for Hinton and Hopfield’s work, the Nobel committee had to stretch to make the argument that their research into neural networks was, in certain ways, inspired by some ideas from statistical physics. This tenuous connection is worth noting for historians of science. But it doesn’t somehow turn computer science, even ideas as lucrative and impactful as AI, into physics.
The task of understanding why the world is the way it is will never end. People will keep finding new questions to ask, and new ways to answer them. Artificial neural networks (or ANNs) will doubtless continue to be applied to new problems in coming years. But their applications in physics have so far been limited. As the physics Nobel citation itself points out, “So far, the most spectacular scientific breakthrough using deep learning ANN methods is the AlphaFold tool for prediction of three-dimensional protein structures.” Maybe one AI Nobel prize this year would have been enough?