8 Facts About AI Research Funding
March 13, 2017, 8 min to read
1// Transparency is sorely lacking
The first thing we noticed when looking for data about AI research funding is that it is incredibly difficult to find that type of information. Research in artificial intelligence is not generally substantial enough to appear as a dedicated line in the budgets of public research agencies - on the contrary of broad academic fields such as “computer science” or “chemistry” for instance.
So the only - but necessarily imperfect - solution is to track programs related to artificial intelligence or its branches over the years. But then finding the relevant programs is a chore. For instance, we once read that the authors of a paper had received a grant from a DARPA (the main R&D branch from the US Department of Defense) program called “Deep Learning”. However, we could not find any evidence of the amount spent on that endeavor, or the names of its beneficiaries - just a technical evaluation report from the US Naval Research Laboratory.
If the public opinion keeps considering the field of artificial intelligence as a key influence on our future - either in a harmful or positive way - then we ought to know what are the areas currently tackled and who is working on them.
2// Funding has not substantially increased - but allocation has evolved
Because of the recent hype of artificial intelligence, we expected to see a significant rise of the corresponding research budget - at least, in the USA. The sum of the budgets from the Information and Intelligence Systems department of the National Science Foundation and the programs related to artificial intelligence from the DARPA has been hovering around $300m and $400m for the last 15 years.
(A figure which is not out of touch with the figure of $1.1bn of US Government investment in “AI-related technologies” in 2015, given by a recent White House report.)
Sources: DARPA and NSF budgets
Since AI research has been around for decades, we guess that far more than the question of the volume of investment, it is its allocation that matters - and has changed. For example, the first mention of “machine learning” in the NSF budget was in 2009. Thus funding probably had its part in fueling the ongoing AI push, that is to say a focus on machine learning and more specifically on deep learning.
3// AI is not an obsession of public funding
The corollary of the previous insight is that artificial intelligence is not (yet) a major area of investment within the immense realm of research. A quick reality check from the NSF:
In the European Union, no section of the €80bn-strong Horizon 2020 research and innovation program is dedicated to artificial intelligence, apart from robotics. Although it could fit well in most of its pillars, such as “Future and Emerging Technologies”, “Industrial Leadership” and “Societal Challenges”. The problem is that the EU R&D policies are negotiated for 7-year periods, and that the current framework was adopted in 2013, when AI had not become a consistent provider of headlines and debates.
4// US military funding is still prevalent
Maybe it was a naive consideration, but we did not expect to find so many acknowledgements of grants from Department of Defense-related agencies (DARPA, Office of Naval Research…) in the papers of researchers located in America, as opposed to the bulk of AI history that unfolded during the Cold War.
The ongoing presence of military goals in the funding of AI research in the US is even stronger when you realize that “securing the national defense” comes third in the mission statement of the civilian National Science Foundation. Which is a striking departure from the equivalent research bodies in Europe. For example, both UK Engineering and Physical Sciences Research Council and French national research policy framework (the “Code of Research”) aim at (i) supporting the advancement of science and (ii) contributing to economic and social progress - let’s stress that their goals are always stated in that order.
Why should it matter? After all, everyone knows that the Internet we gladly use every day is the heir to the military ARPAnet. The answer is twofold.
On the one hand, the mere involvement of the US military in the broad research landscape - beyond areas directly related to defense - can be debated. In this regard, the thoughts that the AI figure Terry Winograd gathered in 1984 are still telling.
On the other hand, whatever one’s view on this involvement and its net effect on our society, its impact on the trajectory of research is definitely not neutral. For instance, the DARPA Grand Challenge, a 2004-05 driverless car competition that was the starting point of all the digital and industrial efforts we see today, was launched with the ultimate goal of automating a third of the US Army’s ground combat vehicles by 2015.
We should never forget the opportunity cost attached to military funding, thus that a bigger importance of civilian goals could yield other interesting directions for the field of AI.
5// China’s influence is growing
Information is scarce, but China is sure bridging the gap with the US. The following graph, taken from a recent White House report, shows that in the area of deep learning, the number of Chinese publications overtook American ones in 2014. America and China are clearly marching ahead and other countries struggle to follow:
Our best proxy for Chinese investment in AI is that numerous teams take part in international challenges and are really competitive. A team from the Chinese Ministry of Public Security won the ImageNet classification challenge in 2016, ahead of a team made up mostly of Facebook AI Research scientists. The year before, Microsoft had won the same challenge - thanks to its Beijing lab.
6// The digital leaders’ strategy is not to fund university labs; it is to compete with them
The giant tech companies of this world do not hesitate to invest in basic research. Their approach is not to crowd out public funding of university labs. For the most part, they have invested in their own research centers. Some of their groups were created in the early 2000s - for instance, machine translation at Microsoft in 2002 - but the effort has increased rapidly in the last few years.
For instance, Facebook hired the deep learning pioneer and NYU professor Yann LeCun to launch the Facebook AI Research lab in 2013, and then established an additional team in Paris in 2015. The Google Brain team dates back to 2011, the firm announced a machine learning group in Zurich in June 2016, and it is currently building a lab in Montreal, a deep learning hotspot.
Some of the most prominent members of Facebook AI Research (Yann LeCun is second from the left)
More and more researchers are joining the GAFA, which are now accused of drying up the talent pipeline of academia. (This is obvious in their direct funding of university labs too, which includes numerous PhD scholarships.) The Uber research strategy in self-driving cars is exemplary of the sometimes contentious relationship between tech companies and academia. Uber signed a partnership with Carnegie Mellon University in 2015, which included a $5.5m gift, but soon the on-demand transportation giant was criticized for poaching dozens of CMU scientists.
7// The most interesting model among research centers is in Germany
The discreet German Research Center for Artificial Intelligence (DFKI in German) was founded in 1988, and provides us with a unique model that could be duplicated in many countries - for its 2015 budget was “only” €41m. It is one of the world’s largest AI labs, with several attractive features:
- As its name suggests, it is solely dedicated to artificial intelligence and its various problem areas - knowledge management, machine translation, robotics… - with close to 500 researchers and hundreds of graduate students.
- DFKI is a non-profit public / private partnership with 3 types of shareholders: 3 German states, 3 German universities and a diverse range of companies - Google, Airbus, Volkswagen, Intel…
- Its 4 locations enable DFKI to set up research groups where they make most sense - e.g. the Robotics Innovation Center is in Bremen, where Mercedes-Benz and Airbus have factories, and the Interactive Textiles group is based out of the more arty city of Berlin.
- DFKI controls the whole chain of innovation, from basic research - 18 research groups - to testing - 6 living labs - and commercialization - 76 spin-offs since its creation.
The Maritime Exploration Hall in Bremen, where aquatic robots are tested under realistic conditions
8// Large companies and governments alike have leverage to take advantage of
Canada is home to two vibrant deep learning research communities, in Toronto and Montreal. The CIFAR (Canadian Institute For Advanced Research) was a key contributor to this dynamism, thanks to a program it funded in 2004, called “Neural Computation and Adaptive Perception”. This initiative gathered the 3 pioneers of deep learning - Geoffrey Hinton at the University of Toronto, Yoshua Bengio at the Université de Montréal and Yann LeCun at New York University. And the annual budget for all of CIFAR’s programs is just $25m.
Given the overall public and private spending on R&D in every country - billions of dollars - this story shows that if we consider AI as a promising research area for our economy and society, then governments and incumbent blue-chip companies could have a tremendous impact on the future of AI by simply redirecting a few dozens of millions of dollars towards this field.
In the meantime, they could contribute to sustain university or public labs as thriving epicenters of AI research, complementary to the substantial efforts led by the leading technology companies.