Related Articles
On Tuesday, the scientists John J. Hopfield and Geoffrey E. Hinton received the Nobel Prize in Physics for discoveries that helped drive the development of artificial neural networks — a technology that is now essential to the operation of search engines like Google and online chatbots like ChatGPT from OpenAI.
The prize was given for a technology that Dr. Hopfield developed in the early 1980s called a Hopfield network and a related technique that Dr. Hinton helped create in the years that followed called a Boltzmann machine. The news surprised many physicists and artificial intelligence experts, including Dr. Hopfield and Dr. Hinton.
In 2019, Dr. Hinton was part of a three-person group that received the Turing Award, often called “the Nobel Prize of computing,” for its work on neural networks. Last year, he made headlines across the world when he left his job as a researcher at Google and warned that the A.I. technologies he helped create could one day destroy humanity.
But he is not a physicist.
He was once introduced at an academic conference as someone who had “failed at physics, dropped out of psychology and then joined a field with no standards at all: artificial intelligence.” Dr. Hinton, a British native known for his dry, self-deprecating humor, enjoyed repeating this story. But he always added a caveat.
“I didn’t fail at physics and drop out of psychology,” he would say. “I failed at psychology and dropped out of physics — which is far more reputable.”
The New York Times reached Dr. Hinton by phone shortly after he learned that he had won the Nobel Prize in Physics.
This interview has been edited and condensed for clarity.
Many congratulations.
Sorry, I can’t talk. I am about to go on with the BBC. Bye.
Hello, again. How was the BBC?
We didn’t connect. I am in a cheap hotel room with no internet.
What was your reaction when you heard this morning’s news?
I was shocked and amazed and flabbergasted. I never expected it.
Neural networks are computer technologies. How does this relate to physics?
Hopfield networks and a further development of them called Boltzmann machines were based on physics. Hopfield nets used an energy function, and the Boltzmann machine used ideas from statistical physics. So that stage in the development of neural networks did depend — a lot — on ideas from physics.
But it was really a different technique — called backpropagation — that was used to build the A.I. models that are used today. That has less to do with physics.
What is the relationship between the Boltzmann machine and backpropagation?
Right now, there isn’t much connection. They were two alternative theories for how we would get neural networks to run.
In the early days, I managed to combine them by using Boltzmann machines to “pretrain” backpropagation networks. But people aren’t doing that anymore.
What do you mean by pretrain?
How long do you have?
Can you explain in language that the readers of The Times would understand?
I am reminded of what the physicist Richard Feynman said when he received the Nobel Prize.
A journalist asked him, “Professor Feynman, can you explain — in just a couple of minutes — what you won the Nobel Prize for?” Feynman apparently replied, “Listen, buddy, if I could explain it in a couple of minutes, it wouldn’t be worth the Nobel Prize.”
Sorry. The BBC is calling again. Bye.
Hello, again. It is safe to say the Boltzmann machine was a dead end for A.I. — that the research went elsewhere?
I think of that idea as something like an enzyme. An enzyme gets you over a barrier — even if it isn’t part of the final solution.
Boltzmann machines were like an enzyme. It got us over the barrier of “How do you train deep neural networks?” It made it easier to train them. And once we had learned how to do that, we didn’t need the Boltzmann machine anymore.
Did you work directly with John Hopfield on any of these ideas?
No. I read his papers. But one of my main collaborators, Terry Sejnowski, worked with Hopfield and did his Ph.D. with Hopfield.
Is it odd that you have received this award for physics?
If there was a Nobel Prize for computer science, our work would clearly be more appropriate for that. But there isn’t one.
That is a great way of putting it.
It is also a hint.
Yes, perhaps we need a Nobel for computer science. In any case, you have won a Nobel for helping to create a technology that you now worry will cause serious danger for humanity. How do you feel about that?
Having the Nobel Prize could mean that people will take me more seriously.
Take you more seriously when you warn of future dangers?
Yes.
<
