We went to Suzhou to find AI’s biggest breakthroughs and bottlenecks

Artificial intelligence experts from Apple, Microsoft, MIT, and more gathered on May 10 at the 2018 Global AI Product Application Expo in China’s rising big data mecca Suzhou.

Aside from the experts, over 1000 artificial intelligence products made by 200 companies hailing from 10 countries and regions demonstrated the power of AI in the slick new business center of this ancient city marked by the famous Suzhou Pants (or as the local officials would prefer us to call it, the Gate to the East).

Here is what speakers had to say about the biggest breakthroughs and bottlenecks in different fields affected by AI technology.

Dr. Guo Di, Machine Learning at Apple, Beijing R&D Center

Smartphones could be a very good platform for AI, said Dr. Guo. Apple is working on smart mobile apps which is a good scenario for AI: first, smartphones are mobile and interconnected and second, the computing power of current phones has exceeded many computers.

Smartphones also have a lot of sensors which can collect data. The difficulties are going to be computing power and data, we have to move from general data to scenario-specific data.

Dr. Guo Di (Image credit: Global AI Product Application Expo 2018)

“For the next generation we would switch from big data to small data, and another possibility is that we do not need data at all. An assistant would just find data automatically. The next generation of technology will be data technology.”

Prof. Polina Golland, Professor of Electrical Engineering and Computer Science at MIT

The most important technology in AI has not been invented yet, said Prof. Golland. Hopefully, in the next few years, algorithms will be able to learn as they go on and bootstrap their learning. Golland also noted that the bottlenecks are the same as they always been with new technology: it’s figuring out what value does it bring to companies.

“From what I see, the technologies are new but the challenges are old. Many companies are excited about using AI but few ask what value does this tech add to consumers.”

Boaz Sacks, Director of Mobileye Aftermarket Division

Mobileye, the autonomous vehicle company that was bought by Intel last year, sees themselves as in the life-saving industry. One of the first AI applications to be commercialized will be collision avoidance and autonomous driving, said Sacks. However, one company cannot do it alone and AI is not enough to achieve autonomous vehicles.

“AI is a derivative of human behavior, what we are lacking is an open transparent standard that defines a common sense of human notions. In our industry, this is: ‘What does it mean to drive safely?’ This is not a commercial problem, it’s a human issue.”

(Image credit: Global AI Product Application Expo 2018)

Prof. Charles Ling, Professor of Computer Science, University of Western Ontario

The major breakthrough during the recent years is deep learning, said Prof. Ling. But the technology is still in trial—we cannot fully explain conclusions reached by neural nets trained by deep learning. That’s why we are still not using deep learning in very important technologies like autonomous driving and nuclear control, he noted.

When it comes to the bottlenecks in AI application, Prof. Ling says that the biggest ones are actually economic and social. He is now working with a startup company on data analysis solution for managing diabetes and obesity. But the bottleneck is that doctors and the pharmaceutical companies in Canada do not want to work with them—for them, healthy people means less business, said Prof. Ling.

Dr. Norbert Gaus, Executive VP, Head of Research and Development in Digitalization and Automation at Siemens AG

AI is not only changing the factory floor but helping us make the industry more efficient. We are now at the process of defining the domain, collecting a lot of different datasets, and accumulating a lot of knowledge, said Dr. Gaus. The next step is taking all this data and describing how it relates to each other as well as putting mechanisms in place that will enable the discovery of links automatically by observing and learning.

“The next step will be automatically discovering links and structures and in the future is what we dare see is learning memories, what we call knowledge graph technologies. This to us is the most exciting and important tech for industrial AI applications.”

Panel at Global AI Product Application Expo 2018 in Suzhou, May 10, 2018 (Image credit: Global AI Product Application Expo 2018)

The biggest bottleneck is that there is a shortage of data which in part can be mitigated with other kinds of knowledge. However, the first practical step is making sure everyone understands what they are doing. AI should be more explainable, not just for the public but for the workers themselves who will have to go through certification programs.

Dr. Chen Tianshi, CEO and Founder of Cambricon

“I think the relationship between hardware and software is like a relationship between water and cups. If you just have a water without a cup you cannot drink it. If you just have a cup without water you have nothing to drink,” said Dr. Chen during the panel. “Without progress in processors we will not see AI but without algorithms, we would not have AI as well.”

Cambricon is a part of China’s new wave of startups working on chip technology with Huawei one of their clients. The crucial task for them are meeting the requirements of the industry of their clients to help them use their microchips better, according to Dr. Chen.

He also believes that aside form deep learning, the next generation of important technology will be related to cognitive technologies—technologies that are able to perform tasks traditionally assumed to require human intelligence: perceiving visual and audio cues, planning, learning, and reasoning.