The Nobel prizes that went to neural networks

Two prizes, one method

The physics prize went to John Hopfield and Geoffrey Hinton for work from the 1980s on artificial neural networks that borrowed its mathematics from physics. Hopfield's 1982 network treated memory retrieval as energy minimization. Hinton's 1983 Boltzmann machine added hidden units and stochastic updates, with an explicit analogy to thermal fluctuations in statistical physics. The chemistry prize went to David Baker for computational protein design and to Demis Hassabis and John Jumper for AlphaFold, which resolved a fifty-year-old problem in predicting protein structure from sequence.

It is the first time we can remember two of the science prizes in the same year going to what is, at bottom, the same family of methods. That is the fact worth sitting with. Whether the committees got the categories right matters less than what it means that a method rather than a discovery is now the thing the sciences want to honour.

The complaint

The complaint arrived fast. Euronews collected the reactions. Jonathan Pritchard, an astrophysicist, wrote that it was "hard to see that this is a physics discovery" and guessed that "the Nobel got hit by AI hype." David Vivancos of MindBigData argued that physics is "tied to something physical" while what a neural network does happens "in the mind of the computer instead of a physical being." Those are two different objections and it is worth separating them.

Pritchard's objection is about what a physics prize is for. If the prize honours understanding of the physical world, Hopfield networks are a strange choice, because they tell you nothing about the world. They tell you about a class of dynamical systems that happens to be useful. Vivancos's objection is more metaphysical and we think weaker. Statistical mechanics has always been about abstractions like ensembles and partition functions, and the Boltzmann machine is a quite literal use of that machinery. If that is not physics, a great deal of physics is not physics.

Hopfield's own line, quoted in the Nature Machine Intelligence editorial, is that "physics is not which kind of problem you are working on, but how you approach a solution." That is a definition by method, and it is the definition the committee implicitly adopted. It is also a definition that would let almost any quantitative field claim almost any other, which is presumably why Pritchard found it unsatisfying.

What the chemists said

The chemistry prize drew less grumbling, and we think the reason is instructive. AlphaFold produced something chemists wanted and could check. Predicted structures were compared against experimentally solved ones, and they matched to a degree that changed what a structural biology lab does on a Monday morning. Baker's methods produced proteins that exist and do things. Nobody argues about whether a designed enzyme is chemistry.

Hassabis, in the Euronews piece, was careful about the division of labour. "The human ingenuity comes in first," he said, "asking the question, developing the hypothesis, and AI systems can't do any of that." That is a modest framing from someone who had just won the prize for the system in question. It is also the framing that makes the chemistry award easy to defend. The instrument was new. The science it served was old.

The instrument view

From where we sit, a small foundation without a wet lab or a large cluster, the most useful way to read both prizes is that the committees decided a learned model can be a scientific instrument, on a par with a telescope or a spectrometer, and that building the instrument can deserve the prize. There is precedent for that. Instruments have won before. What is new is that this instrument is the same object across fields, and the people who built it did not come from those fields.

That has consequences for how work like ours should be judged. If the instrument is what matters, then the questions that matter are instrument questions. How is it calibrated. When does it fail. How do you tell a real prediction from a confident hallucination of a protein fold. These are not glamorous questions and they are the ones a group like ours can actually contribute to, because they do not require having built the model, only having the discipline to test it.

Virginia Dignum of Umeå University told Euronews that "maybe it is time to modernise the Nobel Prizes to recognise that the discoveries that really matter are beyond the traditional division in disciplines." We are sympathetic and also wary. The disciplinary divisions are what give a field its standards of evidence. Chemistry knew how to check AlphaFold. Physics is still working out how to check a claim about a learning system, and that gap, rather than any question of categories, is what we think this month exposed.

What we would want next

The Nature Machine Intelligence editorial reads the prizes as showing "ideas from one field can profoundly impact another." Fine, but the traffic has mostly run one way, from physics into machine learning. We would like to see the reverse tested seriously. Take the diagnostic habits of an experimental physicist, the obsession with systematic error, the insistence on a control, and apply them to the evaluation of learned models with the same rigour physicists apply to a detector. If neural networks are now physics, they should have to meet physics' standards of evidence, and mostly they do not yet.

The other thing we would want is for the next generation of these instruments to be checkable by people who did not build them. AlphaFold earned its prize partly because the structures could be validated by anyone with a crystallography dataset. The models that follow it will deserve the same treatment only if their builders make it possible.