AI Makes Inroads in Physics

But despite its budding ubiquity everywhere else, AI has been a hard sell in physics.

Take Eliu Huerta of the University of Illinois at Urbana-Champaign, for example, who is part of the Laser Interferometer Gravitational-Wave Observatory (LIGO) collaboration. It took Huerta about a year to convince the rest of the collaboration that AI could speed up LIGO’s analysis of gravitational wave candidates “It was a journey,” he told APS News. This February, Huerta and his graduate student, Daniel George, published a paper on their AI-based analysis pipeline.

“People do a bit of naysaying without asking questions,” says Brian Nord of Fermilab, who is part of a team that has used deep neural networks, an AI technique, to identify new astronomical objects in telescope data. AI algorithms demonstrate huge leaps in computational efficiency, but physicists are wary of using them, he says, because their fundamental mechanisms are still largely unclear.

“The skepticism is healthy,” says Nord. “But I think there’s dismissal that comes with the skepticism. I would love for people to ask questions, hard ones. But sometimes … people just say, ‘I don’t believe you.’”

At the 2018 APS April Meeting, several physicists armed with tangible results, including Nord, Huerta, and George, made the case for AI in physics. “It’s harder to dismiss [AI] when you see the benefits it brings,” says Rohan Bhandari, a graduate student at the University of California, Santa Barbara who has developed a deep neural network for analyzing Large Hadron Collider (LHC) data.

Nord’s group is using AI to discover gravitational lenses, massive celestial objects — such as galaxies — whose gravity bends light. These objects leave signature distortions in telescope images that AI can help quickly identify. Understanding those distortions could help answer questions about dark matter, dark energy, and the expansion of the universe.

Neural networks alleviate the tedium of conventional techniques used in the hunt for gravitational lenses: Just a few years ago, “[we] sat in front of our screens, and looked with our eyes through many, many hundreds of square degrees,” said Nord at a press conference on AI in physics research.

Huerta and George have developed a deep neural network to speed up LIGO’s signal identification process. For its first discoveries, LIGO identified gravitational wave candidates using algorithms that match detector signals to hundreds of thousands of “templates” — simulated signals of black hole or neutron star collisions. These algorithms offer a trade-off: You’re more likely to detect a gravitational wave if you compare the signal to as many templates as possible, but more templates take longer to process. More powerful computers could do a better job managing high numbers of templates, but LIGO is already using supercomputers — it’s hard to get much more computational power. “[The community] is really desperate to reduce the number of templates they use,” says Huerta.

So Huerta and George developed a processing pipeline using a neural network that could identify a signal more quickly with less computational power. Instead of comparing signals to templates in real time, the neural network learns the entire library of templates beforehand. “You only need to do the training once,” says George. They found that the neural network-based method was thousands of times faster than template matching.

Soon, says George, they will use the neural network to help LIGO, its European counterpart Virgo, and conventional telescopes collaborate in real time. If LIGO or Virgo can identify and locate a gravitational wave quickly, they can then advise telescopes to observe the same location. These gravitational wave detections can then be paired with the images made with conventional optical telescopes to provide rich physical data about the event in this new era of “multi-messenger” astronomy.

AI can be applied to more than astronomy: Researchers have also begun to use AI to process particle collider data. Bhandari presented his work on a deep neural network for analyzing complicated signals known as jets produced in the LHC. These signals are produced during proton collisions, when constituent quarks and gluons interact via the strong force. Bhandari’s neural network helps to calculate the jet’s momentum, which is used to calibrate the detector.

Fast data processing techniques will be even more necessary in the future, Bhandari pointed out, because they anticipate a massive increase in data from a proposed upgrade of the LHC. “From 2010 to 2017, we collected 230 petabytes of data, and it’s going to keep growing very quickly,” said Bhandari during the press conference. Nord’s field, cosmology, is also expecting a data deluge in the next few years from current cameras — such as the Dark Energy Survey — and new tools, like the Large Synoptic Survey Telescope, which is currently under construction.

Bhandari thinks that physics applications could also help AI researchers understand how the algorithms work. Right now, experts can’t fully explain how the algorithms learn and extrapolate patterns. You can, in principle, write down the equations for the neural network’s operations. But the operations contain so many parameters that it’s difficult to infer what each step is doing. “As physicists, we’re good at looking into black boxes,” says Bhandari. “Detectors are also sort of black boxes. … How do you understand a detector? You do systematic tests to characterize it. Neural networks can be treated in the same way.”

And the black box is becoming greyer: “It has pieces I can pick apart,” says Nord. Researchers have run tests on image recognition neural networks, where they have determined which parts of the algorithm are identifying hard and soft edges in pictures.

Developments in AI for astrophysics observations could easily transfer to other applications, but as Nord pushes for physicists to try these algorithms, he also emphasizes that the work comes with serious ethical responsibilities. “The amazing thing about these cross-cutting technologies is that they apply so generally,” says Nord. “But that’s also the peril of them.”

One plausible peril, says Nord, is misuse of obscure or proprietary AI algorithms by governments. Courts in several U.S. jurisdictions are using AI to predict the risk of future crime in bail and parole decisions, and according to a 2016 ProPublica investigation, the predictions have been biased against blacks compared to whites. Police in Shenzhen, China, use AI-powered facial recognition to publicly shame and fine jaywalkers. And University of Washington researchers have shown how to use AI to make fake videos of President Obama speaking, which indicates that the technology could be used to create fraudulent media for malicious purposes.

---

Brian Nord

Fermilab researchers taught a neural network to distinguish between gravitational lenses and other objects using images (above) as a training set.

---

Nord started using AI in his research partly because he was worried about its potential for misuse. He wanted to educate himself on the technology in order to participate in the policy conversation around it. “If we’re not in the room where the decisions are made, who is going to represent us?” he says.

AI is already everywhere, says Nord. It’s a powerful tool that can help physicists with their research. And in return, maybe physicists can help shape the technology for the good of society.

The author is a freelance writer in Tucson, Arizona.