The 2024 Nobel Prize in Physics goes to the American John Hopfield and the British Geoffrey Hinton. The two AI researchers developed the foundations for artificial neural networks and machine learning, which we now commonly and abbreviatedly refer to as “artificial intelligence”. With their work in the 1980s, the physicists laid the foundations for chatbots that are suitable for everyday use today, such as ChatGPT, but also for special AIs that have long been used in research and medicine to process data.
Artificial intelligence has revolutionized many aspects of our lives and work over the past two decades. We encounter them today, for example, as translation tools or in facial recognition and as text and image generators. Some of these AIs like ChatGPT and Co seem obvious to us today, but their development began around 80 years ago. They were originally developed to process and evaluate large amounts of data in research. In this area, they are actually used today for a wide variety of questions – for example to calculate the structure of molecules. AIs are also used in medical diagnostics, for example to help detect tumors.
Step-by-step development of neural networks
All of these AI systems are based on so-called neural networks – networked systems in which computing nodes are interconnected via connections of different strengths. As early as the 1940s, researchers used the connections between the neurons in our brain, which are more or less densely and intensively connected to one another via synapses, as inspiration for such systems. When we learn, new connections form. Something similar happens with modern artificial neural networks, without them having specific instructions on how to process data: the machine learns independently. Unlike our brain, the artificial neural network does not optimize synapses and functional nerve pathways, but rather signal paths and correlations between input and output.
This year’s Nobel Prize winners in physics pioneered the development of such neural networks and machine learning. “Thanks to their work starting in the 1980s, John Hopfield and Geoffrey Hinton helped lay the foundation for the machine learning revolution that began around 2010,” the Nobel Committee said.
In 1982, the US physicist John Hopfield was the first to develop a – today comparatively simple – network that could process information: the Hopfield network. This is a type of artificial associative memory in which information is stored as energy-efficiently as possible and retrieved later. For the first time, this network was able to compare data patterns with previously stored data sets and associate similar patterns with each other. “The Hopfield network can be used to recover data that contains noise or has been partially deleted,” explains the Nobel Committee.
From the Hopfield network to the Boltzmann machine
The British physicist Geoffrey Hinton further developed this Hopfield network in the 1990s and used statistical methods to build multi-layer neural networks for the first time: so-called Boltzmann machines. Most of today’s AI systems work with multiple levels, which is why they are also referred to as “deep learning” systems. Hinton supplemented Hopfield’s preparatory work with a crucial skill: working on the basis of probabilities according to an equation that the physicist Ludwig Boltzmann had drawn up. Accordingly, depending on the energy level, some states are more likely than others.
Using this principle, Hinton built a neural network that can not only store and process data, but also recognize and interpret characteristic elements in mountains of data. Due to its multi-layer structure, it can recognize not only previously learned patterns, but also unknown patterns based on probability and, for example, describe what can be seen in an image. Today’s artificial intelligences also owe their ability to create new things to Hinton’s model. We encounter them, for example, when a streaming service recommends new series and films that match our current viewing habits, or when online shops recommend products that match our purchasing behavior.
“Through their breakthroughs based on fundamental physics, the two award winners have shown us a completely new way to use computers,” the committee writes. “Thanks to their work, humanity now has a new tool at its disposal. Machine learning based on such neural networks is currently revolutionizing science, technology and our daily lives.”
Source: Nobelprize.org, Royal Swedish Academy of Sciences, Nobel Committee for Physics