The Nobel Prize in Physics 2024 was awarded to John Hopfield and Geoffrey Hinton for their work on neural networks and machine learning.
Frederico Fiúza, a professor at Instituto Superior Técnico and researcher at the Institute for Plasmas and Nuclear Fusion, puts the award into context by describing Hopfield and Hinton’s work and outlining possible futures in the field of machine learning and artificial intelligence, as well as their relationship with Physics.
What is a neural network?
A neural network (natural or virtual) consists of neurons (or nodes) connected to each other which, although very simple individually, can describe highly complex processes as a whole.
The work of these two researchers is related to the development of neural networks, inspired by physical processes and methods from Statistical Physics, which have made a fundamental contribution to the development of machine learning and the way it is revolutionising the most diverse areas, from Computer Science to Medicine and Physics itself.
John Hopfield developed a model of neural networks, also known as Hopfield networks or associative memory, to describe how the brain reconstructs certain memories from partial or noisy input. He showed that the behaviour of these networks resembles how certain properties of atoms (spin) affect each other.
Geoffrey Hinton developed a neuronal network model based on Statistical Physics, which he called the Boltzmann machine (in honour of one of the most important physicists of the 19th century and a pioneer of Statistical Physics), which consists of interconnected neurons capable of making stochastic decisions.
What scientific breakthroughs have resulted from Hopfield and Hinton’s work? What benefits could they bring to the future of machine learning?
These works inspired many scientists to explore neural networks and were very important contributions to the development of state-of-the-art artificial intelligence and machine learning techniques that are revolutionising the way we live and work. Above all, it seems to me that the greatest contribution, that underlies the award of the Nobel Prize in Physics, is the foundation of these techniques in statistical physics. These solid formal foundations in physics (and maths) are critical if we want to understand, gain confidence and make the most of the power of these neural networks to describe highly complex processes. I think this is an important message for the future of machine learning and the benefits it can bring.
In your opinion, how will machine learning impact the near future?
Artificial intelligence and machine learning are already revolutionising our lives because of the way these systems can automate repetitive tasks faster and more accurately than humans. This is already having a big impact on the economy and in many professions, and it will have an even greater impact in the coming years. On the other hand, I am convinced that we are going to see major advances in the ability of these systems to develop more advanced cognitive processes that will have an even more profound (and unpredictable, in terms of both magnitude and time scale) impact in a wide range of fields, such as physics or medicine.
What other advances in the field of physics would you like to have seen awarded with this year’s Nobel Prize?
Several advances deserved to be awarded the Nobel Prize in Physics as well. In fields related to my research in plasma physics, for example, I would like to see the Nobel Prize awarded for the first controlled fusion ignition conducted at the National Ignition Facility in 2022, which released more energy than it consumed. This achievement represents a historic milestone not only in our ability to reproduce the fundamental processes responsible for energy production in the stars, but also to harness these processes to generate a clean, safe, and virtually unlimited source of energy.
How does your work relate to that of the award winners? What is the bridge/relationship between physics and machine learning?
I’m very interested in understanding how machine learning can complement the traditional techniques we use to study and simulate physical systems. Tools such as neural networks are very powerful at analysing big data in complex systems and finding non-linear relationships between them, which can, for example, tell us what laws govern the behaviour of these systems. However, not all of these relationships are necessarily useful. As a physicist, I need these relationships to respect certain principles, such as the conservation of energy. I’m therefore interested in exploring how we can ensure that these machine learning techniques respect physical principles. The ability to guarantee that these tools obey certain principles (or symmetries) could have an important impact on how we interpret and generalise their use in areas other than physics, such as the works awarded by this year’s Nobel Prize in Physics.
Several Técnico professors and researchers also commented on this Nobel Prize in the media:
