Who benefits from the research breakthroughs made in the China-based research labs of American artificial intelligence (AI) companies?
Just five years ago, that question hardly ever came up, and if it was asked, the answer often centered on the shared benefits of global research. Today, much has changed. The field of machine learning has made major strides, China’s technology markets and its surveillance apparatus have boomed, and technological competition has moved to the center of the US-China relationship.
What do all these changes mean for the overseas research labs of leading American technology companies? To answer that question, it’s useful to zoom in on a specific research breakthrough to examine the ideas, institutions, and people involved in it. By tracing those factors over time, a better assessment can be made on where the benefits from this research flow to, and how government policies and corporate practices can best shape those flows.
The case I examine here is the single most-cited paper in AI research over the past five years: Deep Residual Learning for Image Recognition. Often abbreviated as “ResNet,” the 2015 paper is not just the most-cited AI paper based on Google Scholar metrics. With 25,256 citations between 2014 and July 2019, it’s the most-cited paper in any academic field during that time.
The Breakthrough: What’s So Important About ResNet?
ResNet’s central contribution was a technique making it possible to stack many more layers on a neural network—the engines behind much of machine learning today. By stacking more layers (making them “deeper”), the neural network’s performance can be dramatically improved for different tasks: facial recognition, natural language processing, speech recognition, and many other domains.
The technique proved so successful that in 2015, the research team behind the paper took home first place in two of the most important global image recognition contests. By 2017, the technique was one of the core advances behind AlphaGo Zero, the landmark DeepMind system that turned itself into the world’s Go champion by playing against itself.
The Institution: What Lab Produced ResNet?
ResNet was the product of a small research team at Microsoft Research Asia (MSRA), the Beijing lab of the American tech juggernaut.
Founded in 1998 and situated next to Tsinghua University, the lab quickly turned into a research powerhouse, making major contributions to both academic research and Microsoft’s global products. In 2004, MIT Technology Review dubbed it “the world’s hottest computer lab” for its research advances in machine learning, particularly in the areas of natural language processing, speech synthesis, and image recognition.
Research out of the MSRA lab fed directly into software innovations for Windows and advances in graphics for Microsoft’s Xbox, as well as a major improvement in typing input methods for character-based languages like Chinese.
Yet at the same time, MSRA has been perhaps the single most important institution in the birth and growth of the Chinese AI ecosystem over the past two decades. The lab served as a training ground for many future leaders of China’s then-embryonic AI ecosystem, with alumni that include Alibaba’s CTO, Baidu’s President, the head of technology strategy at Bytedance, and the founders of several unicorn AI startups. The Chinese media has even compared MSRA to the “Whampoa Academy of the Chinese internet“—a reference to the legendary military academy that churned out prominent army commanders for both the Kuomintang and the Chinese Communist Party.
The People: Who Was Behind the Breakthrough?
The ResNet paper was authored by four researchers working or interning at MSRA in 2015: Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. (Note: The authors’ given names are presented before their surnames to match how they are presented in international publications.)
All four of the authors got their undergraduate degrees and PhDs at Chinese universities. None of them appears to have lived or worked outside of China before publishing ResNet (Kaiming He did his PhD at the Chinese University of Hong Kong). Both before and after ResNet, the authors have won Best Paper Awards at top conferences and earned first place in global computer vision contests.
And since publishing the paper, all four of them have left MSRA.
Kaiming He, the lead author and most celebrated of the group, joined Facebook AI Research in California in 2016. Xiangyu Zhang and Jian Sun joined Megvii, the Chinese computer vision startup recently added to the US entity list for participation in surveillance in Xinjiang. Sun now serves as the company’s chief scientist. And Shaoqing Ren joined two other MSRA alumni in co-founding Momenta, a Beijing-based autonomous vehicle startup that’s already achieved unicorn status.
To be clear, the fact that ResNet is the most-cited paper does not mean that these are the four most brilliant or important AI researchers of the past five years. Part of what made the paper so widely cited was the fact that it made one critical addition to an already prevalent method, meaning it was relatively easy to replicate in later work. But the researchers behind it made a major contribution to the AI field, and they are putting those skills to work at their new companies both in the United States and China.
Who Benefitted and Who Lost Out?
Returning to our central question—who gains the most from research breakthroughs made at places like MSRA?—the above details provide a whole lot of context but few clear answers.
The MSRA team’s paper marked a major advance for the entire field of AI. Even if other researchers in the field would have eventually stumbled onto similar discoveries, it’s unclear when that would have happened. Since 2015, researchers around the globe have applied the ResNet techniques to AI systems that are both harmful and beneficial.
Microsoft itself also benefited from the paper, though in slightly less dramatic ways than might be imagined. In terms of tangible upsides for business, Microsoft enjoyed some first-mover advantage from developing the technique in-house, but it was likely short-lived. The research team published the ResNet paper online immediately after it won the 2015 ImageNet contest, and it was rapidly adopted by much of the field, including Microsoft’s competitors.
The most substantial gains for the company may have come from bolstering Microsoft’s ability to build a talent pipeline of researchers. The researchers behind ResNet have all moved on, but the work they did there raised MSRA’s credibility with the next generation of computer science graduates.
Beyond Microsoft, Facebook landed the biggest prize when it poached ResNet’s lead author, but Megvii—and the Chinese surveillance apparatus that buys its products—gained big from bringing two of the authors on board. And the Chinese autonomous vehicle ecosystem got a boost from Shaoqing Ren co-founded Momenta.
Alternate Histories and Possible Futures
But any proper assessment of net impact must also examine some counterfactuals. Going all the way back to 1998, it seems clear that if Microsoft had never founded MSRA, that would have significantly delayed the rise of China’s AI ecosystem. Given how primitive the technology was twenty years ago—and China’s reputation at that time as a technological backwater—concerns over China’s future capabilities in AI weren’t on the radar of US policymakers.
But what if those same concerns led US policymakers to force Microsoft to close the lab in, say, 2012?
By that time, two of the papers’ four authors (Kaiming He and Jian Sun) had already been collaborating on research for three years. It’s possible they would have transferred to Microsoft headquarters outside of Seattle and published ResNet there. In that scenario, it’s possible Sun would have continued to work for American companies rather than joining Megvii.
But it’s also possible they would have continued that research at a Chinese institution. In 2013, Baidu founded its Institute of Deep Learning—the first of its kind in China—and it would have been a potential destination for those researchers. If ResNet was developed there (or at a Chinese university lab), it likely would have still been published openly. But what if it had been developed at a government or military-affiliated lab instead, one where open source collaboration is not part of the DNA?
None of these questions yield definitive answers. They can, however, point toward a more productive direction. More granular data need to be gathered on the direction and drivers of talent flows in machine learning. So, too, should there be better understanding of the research cultures at different Chinese institutions. And, hopefully, that combination of data and analysis can help to craft better policies that rely more on the scalpel than the hammer.
Matt Sheehan is a Fellow at MacroPolo. You can find his work on tech policy, AI, and Silicon Valley here.
Get Our Stuff
Get on our mailing list to keep up with our analysis and new products.Subscribe