The 118th Installment
Global AI, Japanese AI
by Hisashi Hayashi,
Associate Professor
The world is in the middle of the third AI wave. I wrote about the AI wave at the end of 2017 when I wrote my last column, but it is not over yet and is likely to continue for some time. I have been studying AI for more than 20 years, and its scope is very broad. There is no doubt that deep learning is at the center of the current wave of AI, but it seems that some people who mistakenly believe that AI = machine learning = deep learning. Machine learning is only one area of AI. Deep learning is a development of one machine learning technology called neural networks. Deep learning has greatly contributed to advances in image recognition and natural language processing, but there are limited fields of application that can be realized by this alone. Neural networks have a long history. However, only a handful of researchers have been studying neural networks for a long time, and many AI researchers began studying deep learning only after the beginning of the third AI wave.
There are several so-called top tier international conferences on AI, and traditionally, IJCAI (to be held in Yokohama in academic year 2020!) and AAAI (internationalized US AI society) are well-known. At these conferences, topics include not only deep learning AI but also other machine learning and non-machine learning AI. In recent years, due to the boom in deep learning, the popularity of NeurIPS, ICLR (mainly for deep learning), and ICML (for machine learning in general) has increased, and has surpassed that of the IJCAI and AAAI. At all of these conferences, presentations on deep learning are very common. Incidentally, IJCAI 2020 was originally scheduled to be held in Nagoya, but the venue was apparently changed to Pacifico Yokohama in order to accommodate the rapidly increasing number of participants.
The author has also attended AAAI 2019 and NeurIPS 2019 to observe, but their vibrancy is amazing. In particular, NeurIPS 2019, where almost all presentations are on deep learning, is extremely popular, and non-presenting attendees must win a lottery in order to attend the conference. About 12,000 people attended the conference, and the venue was so crowded that there were entry restrictions when people tried to enter the poster presentation area. The four rooms for oral presentations were huge, and each could hold several thousand people. And many people gathered each day for the corporate exhibition.
What worried me was the lack of Japanese presence. The overwhelming majority of presentations were from the US and China. In the NeurIPS corporate exhibition, GAFA from the US, DeepMind from the UK, and companies from China were prominent, while SONY was the only Japanese company there. According to what I have heard, employees at many well-known AI companies, such as GAFA and DeepMind, have average salaries in the tens of millions of yen, and they recruit talented deep learning researchers from all over the world through corporate exhibitions at such prestigious international conferences. These companies have written dozens of papers for the conference with a 20% adoption rate and have also written a large number on basic research.
On the other hand, Japanese companies have experienced poor economic performance since before the third AI wave, and many corporate researchers have been required to conduct applied research for quick business applications rather than basic research. During the AI ice age before the third AI wave, AI was considered impractical, so it was difficult to even publicly state within companies that they were researching it. Not surprisingly, Japanese companies have a weak presence at top tier international AI conferences that mainly focus on basic research. Many Japanese companies are unlikely to pay large annual salaries to promote basic research by recruiting leading AI researchers from overseas.
However, in recent years, as more and more Japanese workers have joined US IT firms with higher annual salaries and venture IT firms with comparatively favorable benefits, large Japanese firms have been forced to gradually bring their wage systems closer to international standards. In other words, traditional large Japanese companies also have to change their seniority-based wage systems, reduce the number of middle-aged and older workers with higher wages, and hire talented young people in the AI field at high wages. Recently, I saw a job opening at a national research institute for a fixed-term AI researcher with an annual salary of up to 15 million yen. It looks like even a Japanese venture company that is famous for deep learning pays large salaries to young and talented AI researchers and engineers. If, in the future, Japanese foreign-owned IT companies bring their wages closer to global standards, large Japanese companies will also need to raise their average annual income (or annual income in certain fields, mainly AI) in order to prevent headhunting. Meanwhile, the story will end if the AI wave stops before then. In that case, there would be no reason to pay high wages to AI researchers and engineers, and private AI researchers and engineers who have worked hard in deep learning would be forced to change their fields of expertise.
Returning to the conference, Japan's AI conferences have also attracted a very large number of participants since the third AI wave, and the call for participants for the national conference is closed when the number of participants exceeds the maximum. There was also a corporate exhibition area and research presentations from companies, and Japanese companies seem to be promoting AI research and recruiting AI researchers and engineers. However, although deep learning research presentations are popular, the conferences as a whole also include many AI research presentations other than traditional machine learning. In addition, many of the presentations from companies are not about basic research but applied research. Even a surprisingly high percentage of presentations from university researchers are on applied research. The atmosphere is quite different from that of first-rate international AI conferences, which emphasize basic research and technology itself.
So, what comes next? What is the fate of AI researchers and engineers who were recruited with an emphasis on deep learning and basic research and with high salaries, and of those who are working in Japan’s employment structure that emphasizes conventional AI research other than deep learning and short-term applied research and with low salaries but stability with seniority? Now may be a big turning point.