Artificial intelligence is a powerful tool, but it can also be a “black box” of sorts. What data does AI use to generate its content? What algorithms does it rely on? Much of that information is unavailable to the humans who use AI.
One company is working to change that. Fujitsu is an international communications and technology corporation headquartered in Japan, and it has tasked researchers with building a more user-friendly AI model – researchers like Daniel Quigley, a graduate student at UWM working towards his PhD in linguistics.
Quigley spent his summer in Sendai, Japan, where he and his colleagues worked on causal AI models. Their goal, Quigley said, was to remove the “black box” surrounding AI.
“(Fujitsu) didn’t want to use neural network models because they’re not human interpretable. They’re not human readable,” he explained. “Their model that they’re using for causality is what’s called a ‘glass box.’”
But in the world of AI, computers, and engineering, the expertise of a linguist like Quigley brings a fresh and essential voice.
Linguistics and AI
Quigley’s PhD research centers on formal semantics, or analyzing language meaning using mathematical structures and logical models. It’s a perfect marriage of all of his academic interests – physics, computer science, philosophy, and language. (Quigley earned his bachelor’s degree at UW-Madison and set a record for the number of majors he graduated with: linguistics, anthropology, mathematics, physics, and astronomy.)
His research made him a unique fit for a study abroad program run by UCLA’s Institute for Pure and Applied Mathematics. The Graduate-level Research in Industrial Projects for Students program allowed scholars to partner with corporate sponsors in Japan to work on research projects in different arenas. Quigley applied and was accepted for the partnership with Fujitsu.
With another American student and several Japanese students, Quigley spent eight weeks abroad learning about causal AI and then applying that knowledge to new AI models.
What is causal AI?
“The idea of causality is that you have phrases like, ‘A causes B.’ That’s very different than ‘B correlates with A,’” Quigley explained. “The idea is to have a machine learning model understand and correctly identify that A causes B, and not just A correlates with B.”
It seems simple enough, but it grows complicated quickly.
“(Imagine) you have an arrow pointing from A to B and can say, ‘A causes B.’ I have a probability associated with this causality. But what happens if I introduce another idea or another feature, like C? How does this affect the chain of causal reasoning and the AI?” Quigley mused.
Then they took it a step further. A is a cause and B is the result, but could they start with the B result and conclude that the cause was A? The researchers worked on moving in “both directions,” as Quigley put it. When they visualized these problems in graph structures, he and his colleagues affectionately called them “spaghetti graphs.” All the lines and arrows connecting each node made the whole thing look like a tangled mess of pasta.
Quigley’s expertise was invaluable in deciphering the graphs. He had his background in mathematics, physics, and computer science to help with the technical aspects, of course, but causality is as much a language problem as it is a math problem. He used formal models of causality and tools from graph theory to develop a way to parse the complicated graphs that formed the basis of the AI model.
He also insisted on measures to make sure the AI’s user interface was accessible for all users – things like making sure colors and fonts complied with accessibility standards and that the interface was machine-readable.
Some of Quigley’s work as a linguist was behind-the-scenes, but it was arguably the most important work of all.
“What I brought was something that helped our team work well together: Social science. Linguistics is a social science,” he said. “I had an enthusiasm and an understanding of linguistic or cultural differences between us, as Americans and Japanese working on this team. I already had international experience. I was familiar with linguistic differences as far as what’s expected for communication purposes.”
Real world use
His work paid off: The researchers’ findings were compiled and will later be released for public viewing. Their work could be used to determine any sort of causal relationship between two points, but “What Fujitsu are looking to apply this to are things like medical domains,” he explained. “If I have cancer, for example, what are the causes? What brought this up? Or say I have these symptoms. What will be the results (diagnosis) of this?”
Quigley’s team accomplished a lot during their eight weeks in Japan and their supervisors at Fujitsu were pleased with the work, but Quigley said he would still love to do more. Even so, he enjoyed his time not just in the lab, but exploring Japan as well. He took language and cultural classes and enjoyed his excursions to Japan’s historical sites.
Now, he’s back stateside, but Quigley’s not home yet: He’s embarked on another program with UCLA’s IPAM program researching mathematical intelligences.
After all, if the cause is that Quigley is a talented linguist, the effect is that he’ll find new opportunities wherever he goes.
By Sarah Vickery, College of Letters & Science