Deep Learning Startup PFN Partners With FANUC, Could Save Japanese Manufacturing
January 01, 2017      

In an old building in Otemachi, the heart of Tokyo, is the office of Preferred Networks Inc., a promising deep learning startup. It began when top programmers from the University of Tokyo and Kyoto University met at the International Collegiate Programming contest held by the Association for Computing Machinery.

In 2006, the students founded a company called Preferred Infrastructure (PFI), named after the jargon “purely functional programming language.” The entrepreneurs successfully built a search engine and a big data analytics product without external capital.

In 2014, PFI spun off Preferred Networks (PFN) to focus on building a machine-learning platform for the industrial Internet of Things (IoT) and to accelerate partnerships and fundraising from outside companies. Now, PFN carries traditional Japanese industry’s hopes for the future.

Business takeaways:

  • Deep learning startup Preferred Networks is developing “edge-heavy computing” for big data processing rather than centralized cloud computing.
  • PFN is working with robotics maker FANUC to apply its technology to manufacturing and IoT.
  • The exemplary partnership between the two companies could help Japan maintain leadership in industrial automation.

Products to realize ‘edge-heavy computing’

Machines such as robots, which are effectively mobile sensors, will soon produce a far larger amount of data per year than humans. At that time, it may be not possible to handle all the data in a centralized cloud space. Therefore, Preferred Networks has proposed “edge-heavy computing,” a concept in which decentralized devices process the data, while the devices collaborate with one other.

PFN’s portfolio includes the following products, which cover everything from the raw data-processing tool to the package of solutions for manufacturing customers interested in IoT:

  • SensorBee, a data stream processing engine for IoT
  • Chainer, a flexible and an intuitive deep learning framework
  • DIMo (Deep Intelligence in Motion), a deep learning platform that uses SensorBee and Chainer for edge or fog devices
Deep learning startup PFN's model for applying intelligence to robots and sensors.

Preferred Networks’ DIMo allows for different industries to use deep learning and edge processing. (Click here to enlarge.)

The three industries that Preferred Networks currently supports are manufacturing, transportation, and biotech/healthcare. We’ll focus mainly on manufacturing here.

As for transportation, the deep learning startup has several partnerships with the industry leaders such as Toyota and Panasonic to realize self-driving cars.

For biotechnology and healthcare, Preferred Networks recently founded the PFN Cancer Research Institute (PCRI) and has been working on cancer and genome research with the National Cancer Center of Japan.

In the U.S., most of the big tech players have focused on delivering consumer applications through cloud computing.

By contrast, Preferred Networks has specialized in providing edge-heavy computing for industrial use cases. This may even save “Monodukuri” — the Japanese manufacturing industry. Let’s look at one project example.

Initial encounter with FANUC

FANUC is a market leader in factory automation. The company is well known for its CNC (computerized numerical control) machine and ROBODRILL, which is used in Foxconn’s iPhone manufacturing plant.

In 2015, FANUC invested 900 million yen ($7.3 million at that time) in Preferred Networks. How did this deal — unusual for a typically secretive company — happen?

“It was a coincidence, but it was also destined,” recalled Keigo Kawaai, a business developer at PFN who is in charge of the project with FANUC. “FANUC’s problem consciousness and PFN’s technology matched perfectly.”

Preferred Networks staffers visited FANUC Village in Yamanashi Prefecture (an actual village near Mt. Fuji) to sell PFI’s search engine. They introduced their newborn deep learning startup in the last five minutes of the meeting.

FANUC noticed PFN’s potential and asked the representatives to visit again. When they returned, about 50 people from FANUC side were present to discuss the details of deep learning technology.

FANUC wants to accelerate the adoption of factory automation by adding intelligence and collaboration functions to current stand-alone robots (those without network connections).

In an ideal site, FANUC expects multiple robots to be able to recognize each operating state and share that knowledge with one another.

The manufacturing automation provider was looking for a partner that could “encode the craftsmanship,” and PFN had the right skills.

Equal partnership for deep learning startup

Despite the difference in size and experience, there were no problems between FANUC and Preferred Networks. In August 2015, FANUC acquired a 6 percent share in PFN.

“We are equal partners, not a subcontractor,” said Kawaai. “It’s crucial for us to maintain this relationship, to discuss and try new ideas.”

Current deep learning techniques have been developed since 2012, so large organization probably won’t yet have enough people who are familiar with them. This a great opportunity for deep learning startups, said Kawaai. Both large and small companies should have constant communications to maintain their motivation to work hard and move quickly.

A primary challenge was the difference in technical terms used by PFN and FANUC.

“Engineers had far different backgrounds,” Kawaai said. “At first, we couldn’t understand mechanical or control theory terminology. Neither did FANUC’s engineers; they couldn’t understand networking and machine learning words.”

Regular face-to-face and videoconference meetings helped bridge the gap. FANUC prepared four of its yellow robots for PFN that are still in the laboratory in the Otemachi business district.

FANUC makes its training courses mandatory for all customers, so Preferred Networks staffers attended FANUC Academy.

Moreover, to test the latest robotics and AI technologies, PFN participated in this year’s Amazon Picking Challenge. It earned second prize in picking task and the forth place in stowing task.

Ambitions beyond bin picking

Preferred Networks recently demonstrated the application of deep learning to bin picking, a tedious job for people in factories and warehouses.

In the past, this task was completed by matching image data to prepared CAD (computer-aided design) data and returning the target position to fetch. However, it takes a lot of time and needs expertise for parameter tuning or false detection problems when two objects pile up.

FANUC and PFN let robots learn actively by themselves. If a robot fails to pick up an object, it keeps the image with depth data as a failure, and if succeeds, it keeps it as a success. Then the robot updates the algorithm by using these data sets.

However, the partners have ambitions beyond bin picking. The technology will be applied to much of FANUC’s product line, and the companies are researching other tasks to coordinate.

In April, FANUC collaborated with Cisco, Rockwell Automation, and PFN on its FANUC Intelligent Edge Link and Drive (FIELD) system. An open application programming interface (API) allows more than 200 companies to develop applications for the FIELD system.

Preferred Networks plans to sell the apps it makes with FANUC. By doing so, the deep learning startup can use FANUC’s sales resources and market presence around the world.

The companies are also working on anomaly detection, which can be applied not only to FANUC’s robots but also to other types of machinery, particularly for “zero downtime.”

In addition, the deep learning startup is working with graphics processing unit (GPU) leader Nvidia improve the computing architecture specific to deep learning.

More on Robotics, AI, and Japanese Automation:

Preferred Network’s unique strengths

Google recently published research that found that 14 robotic arms could share knowledge and move faster. While U.S. tech giants want to enter the Japanese market, Preferred Networks isn’t worried about the competition.

“Our uniqueness is the collaboration with the world’s top robot company, which makes products by themselves,” Kawaai said. PFN has a close relationship with manufacturers and can dive into their actual needs and challenges.

“To make it a business, you need to solve the problems which are really painful for the customers,” he said. “We are not sure that how Google thinks about that point.”

The partnership between large enterprises and deep learning startups such as FANUC and Preferred Networks could help Japan remain a player in manufacturing automation.