(Yicai Global) Sept. 1 -- Mathematics can be used to understand the reasons for the success of artificial intelligence applications, and it can also show that some algorithms will not work, which is important for the sector’s development, according to a renowned Chinese recipient of the prestigious Fields Medal.

“It’s wonderful that some AI applications run successfully, but no one has been able to explain the reason yet,” Prof. Yau Shing-Tung said in an interview with Yicai Global. “We’re perhaps still confused about why some succeeded and others failed.”

“Often it’s just thanks to many attempts,” said Yau, who is attending the 2022 World Artificial Intelligence Congress in Shanghai. “Success comes with good luck!”

Yau became famous at the age of 27 for proving the Calabi conjecture, and won the Fields Medal when he was 33, becoming the first Chinese person to win the world’s highest award in mathematics. Yau joined Harvard University as a professor in 1987, and since this April has been teaching at Tsinghua University in Beijing.

Yau offered the example of a current AI application that can easily defeat the top human Go masters, but still seems to perform below par when applied to self-driving. This is an important issue for the development of AI, and mathematics should be a tool to help scientists understand the principles behind it, Yau said.

“We need to know what’s going to happen in the worst-case scenario [for an application] and whether there's any solution to that,” he said. “It takes a lot of data to figure out when a particular AI algorithm will have a problem.”

Not Bothered

In Yau’s view, people do not care about the reasons for the success of AI applications, nor are they willing to work hard on proving when specific applications will fail.

“Because industrial applications are all about success, it seems useless to prove whether there will be failures, and it can be more difficult and requires more effort than the former,” he noted.

But Yau also said that it could be more beneficial to prove that some AI applications are unlikely to succeed.

Yau mentioned Prof. Andrew Yao, a Turing Prize winner at Tsinghua University, who has done a lot of work to prove that some algorithms are unlikely to work, and although his efforts have no direct benefit for developing industrial applications, it can give us a deeper understanding of computing.

Yao’s research conclusions will be very useful in the end, Yao said, but it may take some time before they are completed.

Classic Texts

The AI application that Prof. Yau is most interested in at present is the speech recognition technology developed by iFlytek, he said, adding that the AI language translation technology has made great progress.

Unlike many scholars in other disciplines, mathematicians need to read classic texts, because some of the problems we come across today have been studied by mathematicians two or three hundred years ago, and many of their ideas are still very valuable, Yau said.

He mentioned legendary Swiss mathematician Leonhard Euler, who wrote more than 1,000 papers in Latin, but only two or three hundred have been read by later generations.

“I suggest that iFlytek’s engineers strive for a breakthrough in Latin translation technology,” Yao said “Although it may not have a large market and the commercial returns may be small, I still hope that they can explore classical knowledge.”

Yao added that he had been to Hefei-based iFlytek recently, China's leading AI and speech technology company, to discuss the matter.

“The native cultural influence and communication ability of mathematicians are also very important,” Yau noted. “We must study the papers of first-class European scholars, but Chinese mathematicians have not paid much attention in this respect, and the exchanges with global scholars are not in-depth enough either.

“We can make no progress unless we have lots of exchanges,” Prof. Yau concluded.

Editors: Tang Shihua, Tom Litting