‘We joked: soon AI will also win the Nobel Prize in Literature’

The Nobel Prizes in Physics and Chemistry were recently awarded to research in the field of AI. Artificial intelligence has become an indispensable part of the scientific toolbox, says UG professor Niels Taatgen. ‘Actually, we should teach all our students about the fundamental principles.’

Text: Jelle Posthuma

John Hopfield and Geoffrey Hinton received the Nobel Prize in Physics for their work that laid the foundation for today's artificial intelligence. AI also played a key role in the Nobel Prize in Chemistry. ‘We joked about it here: soon AI will also win the Nobel Prize in Literature,’ says Niels Taatgen, full professor of Artificial Intelligence at the University of Groningen. 

World chess champion

Several news headlines have described Hopfield and Hinton as the founding fathers or ‘godfathers’ of artificial intelligence, due to their research on neural networks from the 1980s onwards. However, the real beginnings of AI can be traced back to the 1950s, says Taatgen. ‘In the early days of AI, researchers followed a logic-based approach; they aimed to create a kind of mathematics of intelligence. The greatest triumph came when a computer defeated the world chess champion in the mid-1990s, marking the peak of this logic-based approach.’

A parallel development was to mimic the brain, more specifically brain cells, to build learning systems, explains the professor. 'As early as the 1950s, a perceptron, a simple brain-based neural network, was developed. Developments in this field seemed promising, but always fizzled out. The conclusion was that these networks were not powerful enough.'

Revolution

In the 1980s, a ‘revival’ in neural networks occurred, thanks in part to the research of Hopfield and Hinton. 'Hopfield and Hinton showed that neural networks could indeed be powerful enough. Hopfield developed the Hopfield networks and Hinton the so-called Boltzmann machine. These networks, which are not widely used today, have a strong foundation in physics, and this was probably an important reason for the Nobel Committee to choose this part of their work. In a way, it was a bridge to the Nobel Prize in Physics.’

After the revolution in artificial intelligence partly sparked by Hopfield and Hinton in the mid-1980s, enthusiasm for AI based on neural networks initially surged but eventually faded again. ‘However, researchers concluded that self-learning systems, also known as machine learning, would be the way to go. In the years that followed, they increasingly turned to alternative approaches, rather than techniques inspired by the human brain.'

Final breakthrough

As of 2010, neural networks were making a comeback. From this point on, sufficient computing power, enough data and new algorithms offered opportunities to work with neural networks. Hinton also played an important role in this ‘revival’, says the professor. In those years, he worked at Google DeepMind, with a strong focus on ‘deep learning’, or artificial neural networks with multiple layers.

However, these learning networks still fit poorly with today's computers, the professor points out. ‘In fact, our computers are based on a 1950s logic-based approach: the architecture stems from mathematics and thus fits poorly with the learning networks. This results in major limitations in terms of computing power and energy consumption.'

Interdisciplinary work

Taatgen is attached to CogniGron, where smart materials are being developed based on the biology of the brain for the latest generation of computers. The researchers from Groningen are working on an architecture better suited to neural networks, aiming to solve problems such as the massive energy consumption. ‘This Prize perfectly reflects the spirit of CogniGron,’ the research hub commented on the X platform after the announcement of the Nobel Prize in Physics. Taatgen: The nice thing is that the Nobel Prize in Physics is awarded to AI. CogniGron is also a collaboration between physics, AI, and computer science, with a focus on materials science.’

The professor sees important similarities with CogniGron in the interdisciplinary way of working, especially in the case of Hinton. ‘We are not only focusing on what we can learn from the brain, but also on what we can learn from human intelligence. This is at the interface of several different disciplines.’ According to Taatgen, generative AI is a big step forward, but still falls short in certain areas. He explains that Large Language Models, a type of generative AI, often make significant mistakes in everyday logical reasoning. Human intelligence is highly flexible and can adapt remarkably well to new circumstances – computers have not yet reached that level.’

Threats and opportunities

Despite this (provisional) gap in adaptability between human and artificial intelligence, Nobel laureate Hinton has frequently warned of the dangers posed by AI in recent years. ‘There are definitely dangers to AI,’ responds Taatgen. ‘Not so much in terms of AI as an overlord – artificial intelligence taking over the world as a ruler, as often depicted in science fiction. But people can, of course, abuse AI. Consider Cambridge Analytica in the US elections. It can actually threaten democracy and privacy.’

At the same time, artificial intelligence has become an essential tool for developments in various scientific fields, Taatgen emphasizes. 'All students need to know the basics of AI. Universities must not focus on Large Language Models like ChatGPT in isolation; they should adopt a broad orientation. Provide students with knowledge on the fundamentals of AI; that is future-proof knowledge.’

All Machine Learning models use the same principle, the professor explains: 'You make a prediction, look at the answer, observe a difference, and then start adjusting various parameters to arrive at the correct answer. If you understand this principle, you will also grasp the new developments in Machine Learning. AI is developing at lightning speed; we don't know where we will be in two years, let alone ten years. This is evident in EU legislation, which is already lagging behind. For that reason, we need to think more fundamentally about artificial intelligence,' the professor concludes.