- AI could be a valuable tool in scientists’ arsenal going forward
- While current AI tools are helping researchers, they aren’t perfect
- We have to understand AI’s limitations if we want to unlock its full potential
When the average person imagines the effect of artificial intelligence (AI), they often think in self-centered terms. They may imagine autonomous cars taking them wherever they go, or losing their job to a machine that can do it more efficiently.
However, scientists look at AI differently. Researchers working in fields ranging from drug development to astronomy are giddy about the possibilities this technology holds. Rick Stevens, associate laboratory director for Computing, Environment and Life Sciences at Argonne National Laboratory, exemplifies this excitement.
Stevens will deliver the keynote at PEARC19, outlining the AI for Science initiative from Argonne. Aurora is slated to be the nation’s first exascale supercomputing system, to be delivered to Argonne in 2021 — and it will be optimized for AI.
“The goal is to take the power of AI and work out how to use that power to advance science and to accelerate science discovery,” says Stevens. “We’re trying to make science go faster by automating the parts that people no longer need to do.”
The problem is that current applications of AI don’t fully fit the needs of researchers. Which means scientists have to rethink what this technology is and what they can expect to get out of it.
AI could fix current problems
While they weren’t developed with specific scientific applications in mind, current AI models work wonders when it comes to some of science’s most repetitive tasks. One example is machine learning in astronomy.
“We're starting to see really interesting progress in interpreting images from telescopes,” says Stevens.
Researchers start with neural networks that were designed for recognizing everyday images, such as pets or houses. One type of network, called a residual neural network (ResNet), can be trained on these basic images to learn how to process visual information and then be retrained for a scientific domain, such as looking at galaxies from a telescope.
“It turns out, that neural network will now be way better than people at classifying galaxies,” says Stevens.
This is one of the main tenants of the AI for Science initiative. Stevens hopes the scientific community can find problems in science “where existing AI and machine learning methods can have an immediate impact.”
“The reason we want to try to find these immediate cases is because we get immediate bang-for-the-buck,” says Stevens. “And you're more or less using off-the-shelf AI methods to immediately get some utility of science.”
By adjusting the AI tools already available, scientists receive the benefits of existing technology without the labor of development.
Learn something, change something
Sadly, preexisting solutions aren’t going to cut it for all scientific applications. Many research projects are too varied or too complex to rely solely on current machine learning tools.
“When you’re building a line of cars, once you build one car, you just tell the machine to make a thousand of them and paint them different colors,” says Stevens. “But in science, you're almost never doing the same thing a million times. You learn something, and then you change what you're doing.”
Science requires deep knowledge of a particular field in order to be able to work within it. Consider, for example, the differences between a physics simulation and a potential AI model.
“We've studied physics for around 400 years, and chemistry for a couple of hundred,” says Stevens. “All that knowledge is sitting in textbooks. It's in people's heads too. But it represents our theoretical understanding of how the world works. It's not in the data.”
“If we want to have AI that knows everything that people know about some scientific area, that AI has to somehow be able to combine what it can learn from data with what it can learn from theories—because theories are what underpins the simulations of the physics,” says Stevens. “And right now we don't know how to do that.”
The broader AI community has spent many years and billions of dollars building generalized tools. Now it’s up to the research community to take the next step and leverage those existing tools to tackle some of science’s biggest challenges.
For example, by working with AI, a small lab with otherwise limited resources could design a new drug that cures a difficult disease or discover a new material that tremendously increases energy efficiency.
“The societies that master this technology and apply it are going to have enormous advantages over those that don’t,” says Stevens. “Instead of worrying about who or what AI will replace, we should be thinking about how we can do more interesting things, more creative things, more powerful things. AI amplifies the power of people.”