The 2024 Nobel Prize in Physics has been awarded to John Hopfield and Geoffrey Hinton “for their fundamental discoveries and inventions that have made machine learning possible using artificial neural networks.”

This decision is special because it highlights a key idea:
Today’s AI “explosion” is not just the result of programming — it is also driven by principles that can be described in the language of physics.

Why were they awarded in physics and not in “computer science”?

Today, many people associate AI only with chatbots, image generators, or recommendation systems. But the Nobel Committee reminds us that one of the core cores of AI is models that find patterns in big data — and for this, approaches well-known in physics are often used:

  • Description of multi-component systems;
  • The idea of ​​energy/stability;
  • Probabilities and statistical laws.

Hopfield’s idea: “Associative memory”

Hopfield’s idea describes a structure that can store and retrieve information.

We’ve all had a similar experience: you remember a word “halfway” — it floats through your mind, but you can’t quite remember it. Then suddenly the correct word “comes to mind.”
A Hopfield network replicates this principle: if you give it an incomplete or slightly distorted template, the system tries to arrive at the “closest possible correct version.”

How?
Hopfield used a well-known idea in physics — energy minimization. The overall state of the network can be thought of as “energy,” and the system naturally goes to where the energy is lowest (the steady state). For example, a ball is dropped across a mountainous landscape and eventually “falls” into the nearest valley. The valley is the “stored word,” and the final stop of the ball is the recovered answer.

This approach is still important today, because it teaches us: a network can “find” the correct answer not by strict rules, but by dynamics and stability.

Hinton’s idea: Boltzmann machine and “hidden” patterns

Hinton was interested in an even more difficult task: Well, let’s say, a network “remembers” a picture — but can it understand what is in this picture? How to “classify” a cat, a dog, a car in the same way that a child learns — without explanations, only by examples?

Here his idea appears: a Boltzmann machine, which is based on one of the main principles of statistical physics — the Boltzmann distribution (probabilities by energy).

An important part of a Boltzmann machine is:

  • visible nodes — where data (e.g., pixels of an image) are entered;
  • hidden nodes — a “hidden layer” that captures features that we cannot directly see;

This “hidden” layer actually means: AI tries to “catch” the rules in the data itself — for example:

  • What makes a “cat” pattern;
  • What makes a “face” structure;
  • What are the similarities between different examples.

The Boltzmann machine is an example of one of the early generative models — that is, it can not only recognize, but also create new, similar patterns.

Why has AI become so powerful in recent years?

The rapid development of AI has been possible due to two factors:

  1. A huge amount of data (for training);
  2. A sharp increase in computing power.

As a result, we got Deep Learning — multilayer, very large neural networks. For comparison, Hopfield’s 1982 network operated with ~30 nodes (a few hundred parameters), while modern large language models can contain more than a trillion parameters. This difference shows us why we have arrived at capabilities like ChatGPT.

How does science use AI today?

AI is not only used in everyday applications. It actively helps physics, for example:

  • Filtering data needed to discover the Higgs boson;
  • Extracting gravitational wave signals from the “noise”;
  • Searching for exoplanets;
  • Predicting the properties of molecules and materials (structure of proteins and materials, new efficient solar cells, etc.).

To view the full text, please visit the link:

https://www.nobelprize.org/uploads/2024/11/popular-physicsprize2024-3.pdf