At the world’s top computer-vision conference last June, Google and Apple sponsored an academic contest that challenged algorithms to make sense of images from twin cameras collected under varied conditions, such as sunny and poor weather. Artificial intelligence software proficient at that task could help the US tech giants with money-making projects such as autonomous cars or augmented reality. But the winner was an institution with very different interests and allegiances: China’s National University of Defense Technology, a top military academy of the People’s Liberation Army.
That anecdote helps illustrate China’s broad ambitions in AI and recent prominence on the field’s frontiers. In 2017, the country’s government announced a new artificial intelligence strategy that aimed to rival the US in the crucial technology by 2020. The latest data on the output of US and Chinese AI researchers suggests China is on track.
Chinese researchers have published more AI research papers than the US for several years, but questions have lingered about the quality and influence of those publications. A new analysis by the Allen Institute for AI shows that China’s share of top AI publications is rapidly approaching that of the US. If current trends continue, the two nations will produce an equal share of top AI publications by 2020.
The Allen Institute analyzed data on more than 2 million AI research publications through the end of 2018 from its Semantic Scholar academic search engine. Comparing US and Chinese AI publications makes it clear that China was an emerging powerhouse of AI research well before the recent national strategy was launched. The country has published more AI papers than the US since 2005, according to Semantic Scholar data.
That trend has long been noted, including by a report on US competitiveness in AI research commissioned by the Obama White House. It has also met with skepticism, because Chinese research institutions have a reputation for low quality and even fraudulent publications.
Yet when the Allen Institute repeated the analysis to include only the research papers cited most often by other scholars, the US did not emerge very far ahead. Extrapolating from data through the end of 2018 suggests China will match the US in its share of the top 10 percent of AI research papers in 2020—the same year China’s government says it wants to draw level with America’s AI prowess.
Citation counts don’t perfectly reflect the quality and influence of ideas, and the Allen Institute is planning more analysis to check to what extent the effect is explained by Chinese authors being more likely to cite fellow nationals. Still, Oren Etzioni, the organization’s CEO, says the findings suggest the US government needs to better support AI research. President Trump recently signed an executive order asking government agencies to do more in AI, but many in the field are skeptical it will be very effective. “It was well intentioned but low on specifics and didn’t deliver the two most important things that we need,” Etzioni says—a more welcoming immigration policy to draw top research talent and significantly more research funding.
Greg Allen, an adjunct senior fellow at think tank the Center for a New American Security, says the Allen Institute analysis should drive home the message that China’s AI ambitions are serious. Lofty bureaucratic strategies and targets like those detailed in China’s AI plan can seem curious when viewed from the US, but they can be effective. “That’s what happens when you call something a national priority and you mean it,” Allen says.
Allen recently published a report on how China’s military and national security apparatus are central to the country’s evolving AI strategy—as evidenced by the military university winning the contest sponsored by Apple and Google. He found the country’s defense ministry is investing deeply in new AI research, for example by setting up two new research centers in Beijing dedicated to AI and unmanned systems. A paper released by one of them in December tried to explain the inner workings of Alphabet’s AlphaZero system that is capable of superhuman performance in both chess and Go.
Data from the Stanford-affiliated AI Index, which tracks the trajectory of AI development using dozens of measures, shows how China’s government is growing its already central role in the country’s research. Government-affiliated AI research papers increased 400 percent between 2007 and 2017, dwarfing the growth from Chinese corporate labs, although China’s state-funded academic institutions still produce most of the country’s research output.
In the US, by contrast, companies such as Alphabet play a much more significant role. US companies contribute nearly seven times the share of American AI publications than Chinese companies do their own countries output.