How artificial intelligence can spur the search for new particles
When hunting for new fundamental particles, physicists have always had to speculate how particles can behave. New machine learning algorithms do not need this.
In a collision that occurred at the Large Hadron Collider this April, individual charged particles (orange lines) and large jets of particles (yellow cones) were discovered. The
Large Hadron Collider (LHC) collides billions of pairs of protons every second. Sometimes this machine manages to rock reality a bit and create something unprecedented in these collisions. But since such events are, by definition, unexpected, physicists do not know what exactly they need to look for. They worry that by sifting through data on billions of these collisions, and by sampling some more feasible amount, they might inadvertently remove evidence of some new physics. “We always worry that we can splash out the baby with water,” says Kyle Kranmer, a specialist in particle physics from New York University, working as part of the ATLAS experiment on the LHC.
Faced with the task of intelligently reducing the amount of data, some physicists are trying to use such machine learning technology as “deep neural networks” to drag the sea of familiar events in search of new physical phenomena.
In a typical case of using this technology, the deep neural network learns to distinguish cats from dogs by studying a pile of photos labeled “cat” and another pile labeled “dog”. But this approach will not work in the search for new particles, since physicists cannot feed the machine images of something that they have never seen. Therefore, they have to do “learning with little supervision,” when machines start learning from known particles and then look for rare events with less detailed information — for example, how often such events can occur in general.
In a paper published in the May online preprints arxiv.org workthree researchers suggested using a similar strategy to expand the bump hunting technique, a classic particle search technique that found the Higgs boson. The general idea, as one of the authors of the work, Ben Nachman , a researcher at the Lawrence Berkeley National Laboratory , writes , to train the machine in search of rare variations in the data set.
Consider the simplest example, in the spirit of the cats and dogs mentioned, as an attempt to find a new animal species in a dataset filled with observations of North American forests. If we assume that new animals will be grouped in certain geographical areas (this idea corresponds to the fact that new particles are grouped around a certain mass), the algorithm should be able to select them by a systematic comparison of neighboring regions. If there are 113 caribou [reindeer in North America] in British Columbia , and 19 in Washington state (despite the presence of millions of squirrels there and there), the program will learn to distinguish caribou from squirrels without directly studying them. “It's not magic, but it looks like it,” Tim Cohen said., a specialist in theoretical particle physics from the University of Oregon, also studying weak surveillance.
For traditional searches in particle physics, unlike the one described, researchers have to make assumptions about how a new phenomenon might look. They create a model of how new particles will behave - for example, a new particle may gravitate toward decay into a certain set of known particles. And only after they determine what they are looking for, they can create a special search strategy. A graduate student usually takes a year of work to do this task, but Nachman believes that it could be done faster and more thoroughly.
The proposed CWoLa algorithm, which means “classification without labels” (MSC), is able to search in existing data for any unknown particles that decay either into two lighter unknown particles of the same type, or into two known particles of the same or different types. Using the usual search methods, it would have taken at least 20 years for teams working on the LHC to sift through all the possibilities that coincided with the second option, and for the first option today there are no search strategies at all. Nachman, who works on the ATLAS project, says KBM is able to do all these searches in one go.
Other experts in experimental particle physics agree that the game may be worth the candle. “We’ve already searched in different predictable places, so it’s quite important for us to go the other way and fill in the voids we haven’t looked for yet,” said Kate Pachal , a physicist looking for collisions of new particles in the ATLAS project. He and his colleagues played with the idea of developing flexible software that could cope with a wide range of particle masses, but none of them were qualified in machine learning. “I think it's time to try it,” she said.
It is hoped that neural networks will be able to detect underlying data correlations that are not available to current models. Other machine learning technologies have already successfully accelerated the effectiveness of certain tasks on the LHC, for example,determination of jets issued by the lower quarks. In that work, it was perfectly clear that physicists miss some signals. “They were missing some information, and if you paid $ 10 billion for the unit, then no information should be missed,” said Daniel Whitson , a particle physics specialist at the University of California, Irvine.
And yet, the area of machine learning is full of cautionary stories about programs that mix hands with dumbbells (or worse)) Some at the LHC are worried that all these short paths will reflect the work of the gremlins in the machine itself, which the experimenters so carefully try not to notice. “When you found the anomaly, it’s not immediately clear - is it a new physics, or is it just something wrong with the detector?” Says Till Eifert , a physicist working in the ATLAS project.