Artificial intelligence is transforming how science gets done
In the wake of the 2010 Deepwater Horizon disaster in the Gulf of Mexico, oceanographer Kaitlin Frasier of the University of California, San Diego, set out to assess the damage that the massive oil spill caused. “We needed to know what happened to marine mammals,” she says.
Specifically, Frasier was concerned with the spill’s impact on dolphin populations. Trying to track the animals from the surface is expensive and time consuming, so Frasier used a different approach: deploying hydrophones to the seabed to passively record every sound in the ocean. By separating out dolphin vocalizations from the general thrum of ocean noise, Frasier hoped to detect trends in the animals’ population density.
The first part of the experiment was successful: Frasier’s hydrophones captured thousands of hours of sea noise that included hundreds of dolphin vocalizations. “We collected terabytes of data,” Frasier says. But that abundance posed a problem. Listening to all those recordings and sorting out the different kinds of vocalizations into categories would leave her no time to do anything else. “I don’t want to spend my life as an expert in dolphin clicks,” she says, “so I thought, ‘I want to automate this.’”
Coming from a family of computer scientists, Frasier was familiar with advances in artificial intelligence technology, so she set up an algorithm to sift through the masses of data and sort the dolphin sounds into categories. “The algorithm is unsupervised, meaning that you’re turning it loose on the dataset,” Frasier says. By the time it finished running, the software had identified seven types of distinct clicks, only one of which had previously been identified as coming from a known species of dolphin.
Humans think analytically. A computer doesn’t have to go through the reasoning process, so it can generate many more hypotheses than a human would.
Frasier’s experience with AI is hardly unique. In recent years, AI and machine learning technology have spread rapidly as a scientific tool, enabling discoveries in fields as diverse as animal behavior, nuclear physics, and exoplanet hunting. Just as the invention of the telescope ushered in heliocentric cosmology and the sequencing of DNA upended our understanding of the tree of life, a technological revolution is spawning a scientific one. But the true scope of AI’s impact may not become apparent for years to come. As its capabilities expand, artificial intelligence may someday change not just how scientists work, but how they think.
The impact of machine learning on scientific inquiry has been magnified by the changing nature of data collection. Long gone are the days when experimenters would collect individual observations and log them by hand. Modern instruments, whether aboard satellites or lurking at the bottom of the ocean, are constantly generating vast amounts of information—so much that human beings are incapable of processing it.
Machine learning algorithms, in contrast, have no trouble sifting through reams of data. They are designed to identify patterns and sort them into categories.
In principle, Frasier could have sifted through all the dolphin recordings by herself, but it would have taken a huge amount of time. And there would have been the issue of replicability. “People are inconsistent,” she says. If other researchers later reanalyzed the dataset, they might group the clicks differently. Feeding it all into an algorithm both saved time and removed the possibility of human bias or error.
It was, though, at least conceivable that a human could have done what Frasier’s AI algorithm did. But that’s not true of every scientific application for which AI is being used.
Remodeling the world with machine learning
Hy Trac is a cosmologist at Carnegie Mellon University in Pittsburgh. He studies galactic clusters, which are giant structures made of stars, gas, and dark matter stretching across millions of light years. A key focus of interest is how massive these clusters are. In the past, the only way to gauge their size was by measuring the relative velocity of the component galaxies and then feeding that data into a statistical model. “You can use the dynamics of the galaxies, how fast they’re moving, to tell you how massive they are using basic physics,” Trac explains.
It’s an imprecise method, however. “We measure tens to hundreds of galaxy velocities, then boil that down to a velocity dispersion, so we’re throwing away a lot of data,” Trac says.
Trac thought he could do better. Working with colleagues at Carnegie Mellon’s School of Computer Science, he created simulated galaxy clusters and then fed data on how they would appear from Earth into a machine learning algorithm. “The neat thing about machine learning is that you can train on all that information, without throwing it all away,” he says.
Not only does machine learning cut the errors in the mass estimate by half, but it promises to provide insight into how that mass is distributed. “We want to make a map of all the matter around a galactic cluster,” Trac says. “The spatial structure is really interesting.”
Unlike Frasier's work categorizing dolphin clicks, these are results no human could achieve even in principle. Applying computational resources in this way opens the door to a level of understanding that would have been impossible otherwise.
No wonder AI technology is rapidly being adopted by other cosmologists. “We’re entering the era of big data,” Trac says. “The number of cosmology papers that feature machine learning continues to rise.”
Someday, AI could not only provide a finer-grained understanding of the data we have but even generate fresh models of how the universe works. That’s right, computers could generate scientific hypotheses on their own. Back in 2009, researchers at Cornell University created a program that could derive simple principles of motion from a dataset that described the motion of pendulums and oscillators. In the future, more powerful algorithms could unearth truly novel theories, effectively automating the expansion of human knowledge.
“Humans think analytically. A computer doesn’t have to go through the reasoning process, so it can generate many more hypotheses than a human would,” says Howard Wactlar, head of the Intelligent Systems division at the National Science Foundation.
And not just more hypotheses, but different sorts of ones. AI’s non-human way of processing information opens the door to insights that strike humans as inscrutable. When Google’s AlphaGo Zero defeated a rival algorithm in a chess tournament last December, its style of play was so unexpected that one human grandmaster likened it to “how it would be if a superior species landed on earth and showed us how they played chess."
It’s easier to achieve this kind of novelty in the rule-bound, unambiguous domain of board games than in the messier domain of the lab, but even here gobsmacking surprises may soon be in store. Eventually, AI may come up with theories that are too complex for humans to comprehend. The idea that computers might rival or even replace scientists might seem a scary prospect to some, but Wactlar believes such a development would be a good thing. “We need more tools to generate ideas,” he says. “We’re not dramatically increasing the number of people in these fields. The more tools we can give them to advance the fundamental science, the more we can accelerate the rate of discovery. I don’t see this as replacing human society but augmenting it.”
AI changing science: Lessons for leaders
- A flood of digital data is transforming science as well as business. Machine learning algorithms help scientists cope.
- Artificial intelligence can not only automate the sorting of data but perceive trends and quantify phenomena that no human could.
- Down the road, AI could generate and test theories of its own, effectively advancing knowledge without human input.
This article/content was written by the individual writer identified and does not necessarily reflect the view of Hewlett Packard Enterprise Company.