2019 Global AI Report: China Leads the World in Number of Research Papers

2019-12-13

2019 Global AI Report covers a wide range of areas, to the extent that its creators (including institutions such as Harvard, Stanford, and OpenAI) have released two new tools simply to filter the information they obtain from it. One tool is used to search for AI research papers, and the other tool is used to research national-level research and investment data.




  • Artificial intelligence research is rapidly advancing


The number of peer-reviewed papers on AI published between 1998 and 2018 has increased by 300%. Attendance at conferences has also skyrocketed; the largest provider, NeurIPS, expects to have 13,500 attendees this year, an 800% increase from 2012.




  • AI education is equally popular


The number of people studying machine learning courses in universities and online continues to increase. It is difficult to summarize the numbers, but a good indicator is that AI is the most popular major among North American computer science graduates. Over 21% of CS PhDs choose to specialize in AI, which is more than twice as popular as the second most popular discipline (security/information assurance).




  • In most indicators, the United States remains the global leader in AI


Although China publishes more AI papers than any other country, papers published in the United States have a greater impact, with US authors' papers being 40% higher than the global average. The United States also allocates the largest amount of funding to private AI investments (slightly lower than China's $12 billion, but higher than the global total of $6.8 billion), and has the highest number of AI patent applications compared to any other country (three times that of the second largest country, Japan).




  • AI algorithms are becoming faster and cheaper to train


Research is meaningless unless it is accessible, so this data point is particularly welcome. The research team notes that the time required to train machine vision algorithms on popular datasets (ImageNet) has decreased from approximately three hours in October 2017 to 88 seconds in July 2019. The cost has also dropped from thousands of dollars to double digits.




  • Autonomous vehicles receive more private investment than any other AI field


Less than 10% of global private investment is allocated to autonomous vehicles, amounting to approximately $7.7 billion. The next highest investments are in medical research and facial recognition (both attracting $4.7 billion), while the fastest-growing industrial AI fields are less glamorous: robotic process automation (invested $1 billion in 2018) and supply chain management (over $500 million).


All of this is impressive, but there is a major caveat: no matter how fast AI advances, it can never match the popularity and hype of mainstream culture and headline news. This may seem odd, but it is worth remembering that despite the thriving world of artificial intelligence, AI itself is still limited in certain important aspects.

The best evidence comes from the timeline of "human performance milestones" in the report, which charts the history of AI reaching or surpassing human-level expertise in various domains.


The timeline begins in the 1990s, when programs first defeated humans in games like chess and checkers, and accelerates with the recent machine learning boom, listing video games and board games conquered by AI. It is mixed with various miscellaneous tasks, such as classifying human skin cancer images in 2017 and performing English-to-Chinese translation in 2018 (many experts have achieved nothing in the latter achievement and pointed out that AI translation still lags behind humans). Despite the impressive list, it should not lead you to believe that superintelligent AI is near.


Firstly, most of these milestones come from the domain of video games and board games, which are particularly suitable for AI training due to their explicit rules and easily simulated features. Such training often relies on AI agents devoting their entire lives to a game and requires hundreds of years of training: this fact highlights the speed at which humans learn compared to computers.


Similarly, each achievement is set within a specific domain. With few exceptions, AI systems trained on one task cannot transfer their learned knowledge to another task. A superhuman StarCraft II bot would lose to a five-year-old playing chess. While AI can accurately detect breast cancer tumors like a radiologist, it cannot do the same for lung cancer (let alone prescribe or make diagnoses). In other words, AI systems are disposable tools, not backup intelligences for humans.


However, there is another aspect that proves AI is not completely useless. As this report shows, despite the limitations of machine learning, it continues to accelerate in terms of funding, interest, and technological achievements.

RECENT AI NEWS

RECENT AI TOOLS