Machine learning wins physics Nobel

If you use AlphaFold, ChatGPT, or other AI tools, you have this year’s physics Nobel laureates, John J. Hopfield of Princeton University and Geoffrey E. Hinton of the University of Toronto to thank. The two men laid the foundation for artificial neural networks. These computer algorithms are designed to mimic the way information is processed in the human brain.

Although artificial neural networks were first introduced in the 1940s, by the 1960s, computational limitations had led to decreasing interest in the approach for real-world applications. That changed when Hopfield published a paper in 1982 describing a dynamic binary model to mimic associative memory, the ability to remember how two pieces of data are connected. For example, while the two words share many letters, a monstera is a leafy green plant and a monster is a terribly frightening creature.

A Hopfield network is a series of binary nodes connected by weighted probability functions. When this network is trained to recognize a specific pattern, those nodes and probabilities change such that when the model is fed a new pattern, it can iteratively course-correct its output based on differences it measures. This ultimately provides an optimized, or “correct,” output which is often the same as the initial pattern. When trained on multiple different patterns, the network can differentiate and sort new data into its correct category. The network is analogous to a brain in that each node-to-node connection becomes a “neuron.”

“The creation and explorations of this network by John Hopfield was a milestone in our understanding of the computational abilities of artificial neural networks,” said Anders Irbäck, a member of the Nobel Committee for Physics.

The next breakthrough in artificial neural networks came when Hinton and his coworkers applied statistical physics to Hopfield’s algorithm, creating what’s called a Boltzmann machine. Rather than focusing on each individual pattern between nodes, this type of model centers on the statistical distribution of those patterns. The algorithm adapts itself based on the number of times it experiences certain patterns so that, with enough training data, a Boltzmann machine can begin to recognize and sort new data into existing patterns –for example, distinguishing between cat and a dog.

To streamline computations in Boltzmann machines, Hinton pruned some of the node-to-node connections making the algorithm less computationally intensive while retaining its ability to learn. These slimmed down versions–called restricted Boltzmann machines–are still widely used today as part of a larger AI algorithm.

The applications of the foundational work by Hopfield and Hinton are vast. Not only do artificial neural networks play a key role in generative AI tools, they are also vital for the kind of complex data analysis done in particle physics, astrophysics, and more. “The impact of AI on chemistry is huge,” says Alán Aspuru-Guzik, a chemical engineer at the University of Toronto. He has used deep learning to identify DDR1 kinase inhibitors and AlphaFold to discover a novel CDK20 small molecule inhibitor.

Although AI is everywhere these days, Hinton was still surprised by his early-morning phone call. “I’m flabbergasted,” he said on a phone call during the announcement, “I had no idea this would happen.”

Even so, Hinton likened the impact of AI to that of the industrial revolution. While machines have long been able to outcompete humans physically, he said that machine learning could one day exceed human mental capabilities. “It's going to be wonderful in many respects,” he said. “But we also have to worry about a number of possible bad consequences, particularly the threat of these things getting out of control.”

Hopfield and Hinton will share the 11 million Swedish kronor (about US $1.1 million) prize.

With additional reporting by Mitch Jacoby and Prachi Patel