Teaching machines to be surprised

The theory that is too good

Physics has a theory of matter called the Standard Model. It is by some distance the most precisely tested idea in the history of science, and in places it agrees with measurement to twelve decimal places. It is also definitely incomplete. It has nothing to say about the dark matter that makes up most of the mass of the universe, nothing about why there is more matter than antimatter, and nothing about gravity. So the theory is both extremely accurate and obviously unfinished.

So something else is out there. The Large Hadron Collider was built, in large part, to find it. Forty million times a second it smashes protons together and watches what comes out. And for more than a decade, everything that has come out has looked exactly like the Standard Model says it should.

Searching without knowing what to look for

Here is the difficulty. The usual way to search is to write down a specific guess: a particular new particle, with a particular mass, decaying in a particular way. You then ask the data whether that exact thing is there. It works beautifully, and physicists have done it thousands of times. But it can only ever find what somebody thought to look for first.

It is a needle in a haystack, except that nobody has told you what the needle looks like.

Looking for anything that does not fit

So the group tries something else. Show a machine a few million perfectly ordinary collisions, without telling it anything about what it is seeing, and it will learn on its own what ordinary looks like, in the way you come to know the sound of your own house well enough to wake up when it changes.

Then show it new data and ask a different question. Not “is the particle I described in here?” but “is there anything in here that does not belong?”

That is the whole idea: teaching machines to be surprised.

The unexpected, systematically

There is a catch, and it is the interesting part. Look at enough data and something will always look strange, for the same reason somebody wins the lottery every week even though nobody is cheating. Ask a thousand questions and roughly one in a thousand will come back odd by pure accident.

This is why physics has such a demanding standard for what counts as a discovery, called five sigma: the result has to be so unlikely to be a fluke, about one in three and a half million, that the field is willing to stake its reputation on it. Run a thousand searches instead of one and the bar has to rise accordingly. Getting that arithmetic right, so that searching more widely does not mean raising more false alarms, is a surprisingly large part of the work.

Every collision leaves a trace

None of this would be possible without the detector: seven thousand tonnes of instrument wrapped around the point where the protons meet, built so that nothing crosses it without leaving a mark. A particle that lives for a trillionth of a second still leaves a trace, and from those traces we reconstruct what happened. We help build that machine too: the innermost layers, closest to the collision, where the tracks are densest and the radiation is worst.

So what does RODEM stand for?

Come and see

Every spring we host the International Masterclasses, where school students spend a day analysing real LHC data next to the people who took it. In summer we bring a cloud chamber to the Nuit de la Science in Geneva, where you can watch cosmic rays, particles that began their journey in an exploding star, draw their own tracks in front of you. Both are free, and everyone is welcome.