This week, TechNode’s translation column gets into the weeds with export controls, with an abridged translation of a deep-dive on AI export controls. TechNode has not independently verified the claims made below. This article was co-translated by Jordan Schneider.
The trade war may (nominally) be at a pause, but the US-China “tech war” is far from over. Over the past several years, both countries have sought to “decouple” from each other, unwinding the complex technology supply chains that bind them together and make them interdependent. The latest salvo from the US was a Jan 6 announcement restricting the export of AI software that can automate geospatial imagery analysis.
This new regulation might sound awfully specific—and it is—but it matters for two reasons. Firstly, it comes on the heels of much buzzed about ongoing review of how the US restricts the export of “emerging technologies.” Despite fears of a sweeping ban on such technologies, this initial restriction has proven remarkably targeted.
As narrow as this ban may have been, Chinese companies are not out of the woods, and it remains unclear what we can expect from the US Department of Commerce. How China’s remote sensing industry reacts to the current export controls gives us a peek into how future export controls may unfold.
Below, an anonymous Chinese analyst argues that while certain industries like self-driving cars may suffer from the export restrictions, the nature of AI research is such that it will not be severely affected unless the US adds considerably more aggressive restrictions.
Swords into plowshares: After the US’s AI software export ban, will an “earthquake” shock China’s remote sensing industry?
By “Tibetan Fox” for Nao Ji Ti, Jan 7.
From the ZTE incident of 2018, to the Huawei saga of 2019, and now to the just-announced AI software export controls of 2020, the increasingly geopolitical international situation has kept our gaze continuously trained on “technologically sensitive” fields.
This time, it is “AI” and “remote sensing” that have been bound together under a ban.
Those friends who keep an eye on current affairs might already know that the new US export controls cover several categories of geospatial imaging software companies, with the fields of UAVs and self-driving vehicles feeling the greatest impact. [Translator’s note: in both English and Mandarin, a variety of terms are used to describe unmanned and autonomous systems, depending on the mode of transport and degree of automation; this article generally translates wurenji as “UAV” or “drone” and zidong jiashi as “self-driving vehicles.”]
In the words of James A. Lewis, a technology expert with a US-based think tank, the Center for Strategic and International Studies, this is “to keep American companies from helping the Chinese make better AI products that can help their military.”
But wait. How is it that artificial intelligence and spatial remote sensing technology are relevant to military matters?
Not long ago, the US demonstrated for the world a “new game mode” for AI-powered remote sensing—the precise “decapitation” of Iranian regional top commander Qasem Soleimani, with his executioner not even being a person, but a Reaper UAV. From a thousand miles away, the drone’s operator lobbed four Hellfire missiles down at him, and boom…
Of course, in China—and an overwhelming majority of countries—AI and remote sensing are more connected with applications such as agriculture, surveying, prospecting, mapping software, and so on. In that case, will these software export controls bring about a “semiconductor-style” crisis [i.e., a “ZTE moment”] for these emerging fields?
Swords into plowshares: the epochal link between AI and remote sensing
Viewed as a whole, the application of AI to the field of remote sensing builds on three core functions:
- High-intensity, real-time data processing that can handle diverse data sources and structures
- Effective obstacle avoidance and automated operations that can respond to complex weather and environmental conditions
- Improving environmental monitoring, decision-making, and early warning capabilities in areas that humans cannot reach
Artificial intelligence adds value in one more way to the domain of remote sensing: namely, by combining UAVs, self-driving cars, and other sensing devices, one can endow endpoint devices with an intelligent “brain,” substituting for humans to complete tasks that would previously have been unachievable.
The AI ban: will an ‘earthquake’ shock Chinese remote sensing industries?
Just how much of a “shock effect” will US export controls have on Chinese remote sensing activities?
Looking at where things stand, it seems as if everybody has finished watching the show and snacking on melons, each wandered home, and got back to whatever each was supposed to be doing. Since nobody has had to whip out a “spare tire” following the imposition of export controls, unlike with the semiconductor industry, no one is now shouting themselves hoarse with criticisms or questions.
Is the role of AI software in remote sensing not big enough?
First, the export controls are on software that automates the analysis of geospatial imagery and will decidedly not cause a quick chain reaction.
The main function of this type of software is in training deep convolutional neural networks to automatically analyze geospatial imagery and point clouds. For example, discerning between vehicles, houses, and other such objects, or reducing pixel variation during scaling, resizing, and other operations.
Overall, when it comes to these AI software restrictions, there just isn’t enough for everyone to start panicking.
On the one hand, the export controls will trigger a relatively long-lasting tug-of-war, and will also implicate a relatively large number of interest groups. These export controls will impact several industries that use the related software as a base to develop aerial photography, 3D maps, and so on, because many of the software products in question are built on platforms such as AWS and GCP that directly provide map processing APIs.
But products built on open-source software like TensorFlow and PyTorch will in fact not be impacted, and affected companies, platforms, and communities can take active measures to avoid this problem.
The most direct example comes from October 2018, when the US Department of Commerce added eight Chinese companies, including Hikvision, Dahua, iFlytek, Megvii, SenseTime, Yitu, and others to its Entity List, stipulating that NVIDIA, Intel, and other companies were not to sell microprocessors to them. But to this day, China remains an important and indispensable market for these US enterprises.
This is not the first time that the US has restricted the export of certain technologies, and it will certainly not be the last. Everyone has become used to barrier after barrier, and their own industries have begun to grow… All one can say is, there is real truth to the name “Chuan Jianguo”. [“Chuan Jianguo” is a nickname for Trump among Chinese Internet users, translating roughly as “Trump Builds the Country.” It is a sarcastic compliment suggesting he is helping China, particularly by engaging in a trade war that forced China to reduce US dependency.]
Of course, the more important thing is that this software “blockade” could, from an objective point of view, create a lag in the development of China’s UAV and self-driving car industries. On the one hand, there are numerous prerequisites to combining spatial remote sensing software and artificial intelligence, such as the practicalities of integrating high-resolution remote sensing image wavebands, specialized microprocessors and datasets for vertical fields, and so on, all of which will impact the accuracy and usability of AI in the domain of remote sensing.
At the same time, China itself has sufficiently strong prior experience and R&D superiority in the field of AI algorithms, with examples such as DJI, the UAV company with the best techniques; Huawei, which has invested much energy into AI edge computing; Baidu, with its R&D capabilities in self-driving vehicles; SenseTime, which has pushed out an artificial intelligence analytic platform for remote sensing; and so on. With all this, export controls should make one less tense.
The most crucial aspect is that the US’s AI software export controls could well “kill 800 invaders at the cost of 1,000 troops.” As everyone knows, the rapid development of artificial intelligence is inseparable from an open atmosphere and industrial environment, and software development in particular relies on open-source, trust, and a global communication environment.
Much of US industry and academia also need China’s strengths to develop projects together, which is why when the US requested that GitHub stop Chinese registrations, GitHub considered opening up a subsidiary in China. [In fact, although GitHub has restricted the accounts of developers from countries under US trade sanctions including Iran, Syria, and Crimea, Chinese users are not affected, nor does it appear that registrations from the affected countries have been halted. However, GitHub appears to still be actively considering opening a subsidiary in China over related concerns.]
At the end of the day, China is an enormous market for AI R&D, applications, and supply chain manufacturing; being “invisible” to China also means China will be “invisible” to onesself, which will only accelerate the birth of Chinese versions of Android, iOS, and Github.
This is also why some web users have suggested that this round of export bans on geospatial software is similar to the 2000s’ “strong cryptography export ban,” in that it is doomed to fail: You can stop a supplier with unique technologies from providing restricted hardware components, but you cannot halt knowledge transmission in an entire field!
Indeed, this time the “knowledge” is already in China, and does not even need to be “exported!”
AI remote sensing: entering ‘No Man’s Land’ as the next step
Of course, this does not imply that China’s intelligent remote sensing technology industry can rest completely at peace.
Even though a particular round of software export controls may not cause serious injury to the entire industry, when viewed as a whole, the field of remote sensing still has many more areas that are not mature.
For example, in the domain of China’s remote sensing, AI inference algorithms have already made considerable strides, but specialized inference chips are still being behind, with most algorithms still employing general-purpose microprocessors produced by NVIDIA and others. As a result, certain gaps in efficiency may emerge when using domain-specific architectures (DSA). Also, the precision level of manufacturing needs to be improved, which affects R&D of top-notch remote sensing equipment similar to the Reaper UAV.
Additionally, having previously raised the importance of data to intelligent remote sensing, it is worth noting that much geospatial remote sensing data is provided by coordination with satellite networks. This creates a particular need for aviation information industries to build systems that can guarantee high spatial resolution, high temporal resolution, high spectrum resolution, and high data quality; for example, the formation of the recent Beidou satellite network has valuable and long-term significance for intelligent remote sensing. And higher-quality, higher-complexity remote sensing data will also require greater computing power to process, further requiring that national semiconductor industries continue to forge on.
Furthermore, remote sensing technology also relies on comprehensive technological upgrades in sensors, monitoring equipment, UAVs, and other fields. For example, the MQ-9 Reaper is “nearly soundless” during flight, and only because of this can it attack its target without being found out before the fact. Looking at the current situation, no matter whether it is the application of satellite remote sensing imagery or the technological threshold of artificial intelligence, all lack adequate relevant professional support, and there remains a need to improve the iterative speed of the integrated process from image collection to understanding to analysis to training and beyond.
When all is said and done, as remote sensing gets more and more developed, and lone warfighting units become more and more integrated with AI, in time we will still see many more creations that bring prosperity to society and humankind. Against this trend, what is needed is not the blind optimism of motivational catchphrases, only that we band together in strength as we face great hardship in breaking new ground.