Nobel sparks debate as AI pioneers honoured for neural nets, not physics

Honoured for neural nets, Hopfield and Hinton's Nobel raises questions on awarding contributions outside traditional physics

The 2024 Nobel Prize for physics has been awarded to two living legends of computer science. John Hopfield and Geoffrey Hinton were pioneers in conceptualising neural nets. They separately wrote seminal papers that led to the establishment of the field of machine learning (ML), which is foundational to the development of artificial intelligence (AI). However, as even the prize citation makes clear, while the two awardees used tools adapted from physics to create neural nets and drive research in ML, they did not directly make contributions to the discipline of physics. This is why the award has proved somewhat controversial with several working scientists pointing out that a Nobel for physics may not have been an entirely appropriate way to acknowledge the massive contributions of this duo.

In the last two years, the advent of large language models (LLM) such as ChatGPT and all its competitors has led to an explosive proliferation of AI-driven tools. AI used to be considered an obscure area of research, except by experts in the domain. It is now being deployed across multiple industries and has thus entered the lives of billions of people who know little or nothing about the ways in which AI works under the hood. It is even true that AI is contributing to research in many scientific disciplines, including physics. In that sense, it can be argued that the laureates are enablers of physics research. The Nobel Committee decided to recognise the contribution of physics to the rapid development of AI.

A neural network can be loosely described as a computer network, which is organised to store and process information in the same way as a biological brain. A net has nodes, and layers, connected to one another in a model that mimics the way neurons connect up a physical brain. A neural net sets weights to information in the same way that human brains prioritises information as important, urgent, or less consequential. Neural nets “learn” using training data in the same way the brain does. Nets store patterns and recognise them in the same way the brain does, and the nodes pass on information to connected nodes when the weight is higher than a chosen threshold.

Dr Hopfield, 91, wrote a key paper in 1982 where he created this basic model of a neural net, with the concept of weights and thresholds for output/input. In 1985, Dr Hinton, 76, adapted a model from statistical mechanics —  the mathematical tools used to study the behaviour of gases — to create a more refined model for operating neural nets. From there to the state-of-the-art in AI has been a long journey but one that is still recognisably influenced by those papers. Modern AI utilises much more powerful chips but it uses a conceptually similar framework. And, of course, modern AI, with its huge number-crunching capacity, is used to detect patterns in scientific data, such as finding new exoplanets by sifting astronomical observations, or finding better ways to control nuclear reactions. Modern science is interdisciplinary, so a case can be made for flexibility in this regard. Modern science is also based on teamwork, and breakthroughs (as in AI) depend usually on contributions from multiple individuals and many teams. Perhaps the Nobel Committees could consider altering the requirement that a prize cannot have more than three awardees.