Why We Can’t Build an IoT with Computer Science Alone?
The real essence of IoT lies in integrating fog computing with device development, says PFN founder Toru Nishikawa
by Nobi Oda
“The Internet of Things” (IoT) refers to a network scenario that connects different devices to the internet and strives to automate, optimize, and increase the efficiency of the data exchange between them. So far, AI has played a huge role in the growth of this new field, which suddenly finds itself at the center of the tech-world spotlight.
Today, a new venture started by students at the University of Tokyo has begun utilizing AI developed for search engines to push the IoT concept another step forward. Toru Nishikawa, the young leader of PFN (Preferred Networks), has had his eye on a serious foray into the IoT industry for a couple of years. In March 2014, he formally established PFN as a spinoff of his original company PFI (Preferred Infrastructure).
Nishikawa explains the reasoning behind the split this way:
“In the last few years, IoT — or more specifically, AI applications for IoT — has been expanding rapidly. At PFI we’d accumulated a lot of knowledge that allowed us to create deep learning technologies and other things like it. But competition in that field is intense, and if you can’t raise your operating speed your development projects won’t get anywhere. At PFI, we made a rule of not accepting external capital and not working with subcontracted or commissioned developers, but at PFN I decided I wanted to cooperate more proactively with outside companies, to help lead the way in developing new technologies.”
At PFI, Nishikawa and his team devoted all their energy to creating high-quality programming technologies, driven by a sort of techie idealism. But at PFN, they’ve begun partnering with the companies that actually have to implement their software, and Nishikawa says he is now pursuing development strategies that are more conscious of how software interacts with the devices that run it.
“I first heard about IoT around three years ago, during a visit to the Cisco home office where I learned the principles of fog computing and had an exciting exchange with one of the supervisors there. At the time, I never imagined that IoT would receive the kind of global recognition it’s getting today.”
In today’s computing environment, where the networkability of devices grows explosively every year, cloud computing has become a mainstream technology. By 2020, it is estimated that more than fifty billion terminals will be connected to networks.
The IoT craze grew out of that reality as more and more devices began connecting to clouds. As innumerable devices began exchanging data via cloud computing, engineers became more aware of the need to optimize their own operations for maximum efficiency in response to network environments that change every second.
However, streamlining these processes is extremely difficult with just a device and a cloud. The distance between the two is too great to allow for instantaneous responsiveness — and the devices, in particular, are so scattered geographically that they’re incapable of processing all the relevant information in real time.
In response, Cisco introduced a concept called “fog computing.” Basically, fog computing works to enable real-time information handling and strives to connect devices to a smarter network.
“At PFN we had a similar concept, which we called Edge-Heavy Computing,” Nishikawa explains. “We saw that Cisco’s thinking had a lot in common with Jubatus, one of our projects, so we decided to partner with them and see what we could do.”
This turn towards network efficiency is in line with a trend seen today in the “Industry 4.0” movement first initiated by the German government. There, German policymakers hoped to create a revolution in production by building IT infrastructures into the manufacturing industry in order to create more intelligent, more efficiently networked production processes.
“With Industry 4.0,” Nishikawa says, “the idea is to gather data into the cloud so you can analyze it and put it to use. On our end, we focus on improving the network’s real-time responsiveness.”
IoT has potential applications in so many fields that it isn’t possible to name them all, but PFN has concentrated in the areas of automotive manufacturing and industry robotics.
“If you’re trying to build an automated driving system for a car, what you’re essentially doing is trying to synthesize the operations of a huge number of machines. You have to create a driving system that is safe and comfortable, and that can keep all of those machines in sync while remaining responsive to pedestrians and other cars in the street, traffic laws, and the condition of the pavement. Or, if you’re talking about assembly robots, you have a situation where a huge number of robots are working together to build a single product. If you can create an environment where these robots can share the information and knowledge they gain without any lag — an environment where, say, one robot can fix another when it breaks — you’ll be able to increase production speed and volume significantly. This is exactly the kind of thing our deep learning technology was designed for.”
If these ideas bear fruit, fully unmanned assembly plants and autopilot for cars may become realities. IoT machines will be able to learn and gain experience from each other — and as time passes, we will enter a world of endlessly increasing precision, efficiency, and safety.
Nishikawa identifies two major problems stand between PFN and that dream: latency and data volume.
“Latency” simply refers to the time lag between when information is sent and when its effects are felt — in other words, the time between when a device sends data to the cloud and when its results are sent back. If this lag lasts even a tenth of a second, automated systems will stop working. If you think about what this would cause in a car on a freeway, you’ll begin to see how indispensable real-time responsiveness between clouds and devices can be.
The data volume is visible even at the level of individual machines, because every single device can process data so minutely detailed it might as well be infinite to us. A machine that, say, processes 200 kilobytes of data in a millisecond will process 200 megabytes every second — and if you have 100 machines in an assembly line, that means they collectively process 20 gigabytes of data every second they spend working. No computer processor can keep up with this.
Because of these limitations, the question of how to improve network functionality, and how to implement data processing at the network level, has become more and more important.
In addition to its work with industrial robotics and the automotive sector, PFN has begun collaborative research with the Kyoto University Center for iPS Cell Research and Application, operated by Nobel Prize-winning cell biologist Shinya Nakayama.
“Right now, this research is mainly confined to the life sciences, but there is a lot of speculation that it could become automated and expand into mass production. High-volume data collection has already become a reality thanks to the mechanization of research. In the future, I think data-driven drug development and deep analysis of side effects will make it possible to develop order-made medications optimized for each patient.”
AI logic functions are advancing quickly as well, and PFN has begun applying insights gained from the latest AI research to network functionality. “It’s hard to achieve real breakthroughs just by making technical improvements,” Nishikawa explains. “You have to improve functionality and flexibility on the device end too.” On top of that, he says, you have to consider how to work existing hardware into the network structure.
And it is in that area that Nishikawa believes Japan’s advanced “maker skills” can have the biggest impact. In their work on industrial robots, PFN has partnered with FANUC, a global company that builds many different kinds of industrial robots. By combining FANUC’s control technology with PFN’s machine learning technologies, Nishikawa hopes to push manufacturing another step forward. You might think of this as taking “integral production” — one of the Japanese manufacturing sector’s signature strengths — to the next level.
In Japan, manufacturers have achieved unparalleled functionality by intricately synchronizing the operations of the basic technologies they use. However, as electrical appliances have become more and more universal, so-called “module development” has spread to the rest of the world. By standardizing isolated technologies and treating them as pats of an integrated circuit, Japanese manufacturers took a global lead in developing low-cost, high-functionality appliances. But because this mode of production did not require sophisticated training, it was easy to introduce in developing nations, and Japan gradually lost its worldwide supremacy in the area.
In listening to Nishikawa’s story, however, it starts to feel as though a new kind of integral production, made possible by the shrinking gap between networks and devices and the machine learning it enables, may be just what’s needed to advance manufacturing to the next level.
“The know-how behind something like this can’t be created in a day and a night. Japan’s greatest strength has always been its obsessive pursuit of high quality and its refusal to neglect minor details in that pursuit. I think this perfectionism will become a greater and greater asset for us as time goes on. In other words, I think the world may be nearing a paradigm shift where you can no longer take over the IT world with a single idea.”
In human terms, this is amounts to saying that the body is becoming as important as the brain. In the field of raw intelligence, algorithmic processing overtook our frontal lobes a long time ago. The most famous example of this are computers that can defeat the best chess players in the world.
“But in human functioning, embodied sensation and perception play a huge role in how we process the world, and you can’t say the tech world is anywhere near as strong in this area as it is in data processing. Precisely for this reason, it’s important to keep creating better devices at the same time that you build better AI.”
Today, PFN is structured like a think tank, with about thirty employees working in offices in Tokyo and San Mateo. But this small company may hold in its hands the key to the future of manufacturing.
(Photograph: Ryosuke Iwamoto, Translation: Michael Craig)
Originally published at ignition.co.
