This Startup Just Built Japan’s Most Powerful Supercomputer

Preferred Networks is making changes in Japan.

Over the past few years, this AI startup has raised more than $130M in venture funding and grown to more than 130 people.

If you live outside of Japan, you might not have heard of this team, but they are working with Toyota to create the next generation of driverless cars. They are working with Japan's most advanced industrial robot manufacturers to improve efficiency. They are also working with many financial institutions on fraud detection.

Oh yes, and they also built Japan's most powerful commercial supercomputer.

Today we sit down and talk with Daisuke Okanohara, the technical co-founder of Preferred Networks. Daisuke and I talk about the story behind Preferred Networks, he also shares his challenges and current strategies for maintaining the company's experimental and engineering culture as it grows larger and more structured.

Daisuke also talks about his time at Google, how Japanese AI stacks up to China and the US, and why he’s convinced that their biggest competition is going to come from somewhere you would never expect.

It's a great discussion, and I think you'll enjoy it.

Show Notes

What edge-heavy computing is and why it's important How a Google Internship changed Daisuke's outlook on AI The future of driverless cars at Toyota Why the team decided to build Japan's most powerful supercomputer Why you can't sell disruptive products to large companies How to keep a curious spirit even as your company grows Where the real competition in AI will come from

Links from the Founder

Everything you ever wanted to know about Preferred Networks

Check out their Homepage Follow them on Twitter @PreferredNet

Check out Chainer Preferred Networks free open source AI library

The core Chainer project PaintsChainer Cupy Chainer

[shareaholic app="share_buttons" id="7994466"] Leave a comment Transcript Welcome to Disrupting Japan, straight talk from Japan's most successful entrepreneurs. I'm Tim Romero and thanks for joining me.

Preferred Networks is without question the brightest star in the constellation of Japanese AI startups. It attracted about 130 million in venture funding and have grown to more than 130 people over the past few years.

Of course, if you don't follow AI, you might not have heard about them at all but they are the technology behind Toyota’s driverless cars, some of FANUC’s industrial robots, many cutting-edge applications in other verticals, and as a side project, they also built Japan's most powerful commercial supercomputer.

It's an interesting team to say the least and today, we sit down and talk with Daisuke Okanohara, Preferred Networks’ technical cofounder.

We talk about how Preferred Networks got started and got to scale and he also shares his challenges and strategies of trying to maintain the company's experimental and engineering culture as it grows larger and monthly revenue pressures increase. Daisuke also talks about his time at Google, how Japanese AI stacks up to China and the US, and why he's convinced that their biggest competition is going to come from somewhere you would never expect it.

But you know, Daisuke tells that story much better than I can, so let’s gets right to the interview.

[pro_ad_display_adzone id="1411" info_text="Sponsored by" font_color="grey" ]

[Interview]

Tim: So I'm sitting here with Daisuke Okanohara, the cofounder and Executive Vice President of Preferred Networks, Japan's leading and probably most innovative AI startup.

So thanks for sitting down with me today.

Daisuke: Thank you very much.

Tim: So Preferred Networks talks a lot about the importance of edge -heavy computing. So can you explain exactly what edge-heavy computing is and why it's important?

Daisuke: Cloud computing is one of the most important trends in the IT area and most people believe that most computations or operations should be done at a data center or across site, and it's okay if we deal with fragile information but when it comes to solving real-world problems like operating robots or autonomous driving, we need to process data at the edge site or near to the device.

Tim: So just so make sure I understand it correctly, edge-heavy computing is important because of the latency of how quickly?

Daisuke: Yeah, and reliability.

Tim: Ah, right, right, because you want to always have connectivity?

Daisuke: Yes, current internet, it's not reliable to use for the mission-critical tasks.

Tim: Okay, so it's really a trade-off between the latency versus the amount of computing power you have? So if you can wait for the results, it's great to compete in the cloud but the closer to real-time, the more important edge computing is?

Daisuke: Yes.

Tim: Alright, that makes a lot of sense. But isn't the promise of big data that you need like, huge data sets? So when you are using edge-heavy computing, do you still send all the information to the cloud for analyzing later or do you generally try to make self-contained systems?

Daisuke: Our goal is to extract the information at the edge site and only send the essential information to the data center or other place, so we do not even know the data center or cloud but when there are many devices, we cannot send all the data to the cloud and we need to process most of the computation at the edge site.

Tim: Okay, so it's just a more clear separation of training data and execution data?

Daisuke: Yeah. In the current status, training requires much more computation part so we need to train the model at the cloud site and the operation execution or inference at the edge site.

Tim: Okay. Now, Preferred Networks is involved in AI applications ranging from like, automotive and factory automation, life sciences, network security, and I want to talk about all of those in a little bit but before that, I want to talk a little bit about you.

So before you got your PhD, you were an intern at Google in San Francisco? How did that happen?

Daisuke: Yes. At the time, I started Natural Language Processing and many of my friends started working at Google. I was also interested in how Google solves a problem and how the people in Google are working, so that they could produce many excellent products.

Tim: So did you apply to Google in San Francisco from Japan or did you apply to Google Japan and said, "You need to work with this group in San Francisco"?

Daisuke: Actually, I applied to Google Japan but at the time, Google Japan did not have enough resources to accept intern members, so internship students went to San Francisco.

Tim: Okay, so it sounds like you had a real passion for AI before you started working with Google. So what did you take away from that internship?

Daisuke: Before working at Google, I did not imagine how AI can be used for solving many problems. For example, the search engine, motion transition, image recognition, speech recognition, and so on. So many products and services use machine learning as an essential tool.

Tim: That's interesting because I think in a lot of technologies, Japan is very strong in academic research but tends to be weaker in creating new products and bringing new products to market, not just AI.

Daisuke: Yeah, I think Japanese is a bit conservative and they hesitate to do different things. In my opinion, it comes from Japan is a monoculture. All people speak Japanese and spend the same expenses wear the same clothes and maybe in the companies, so it is very tough to start new things. I think that diversity is very important to bring new products to our futures.

Tim: Yeah. It makes sense working with Google, you understood the importance of practical applications of AI but then when you came back to Japan, you went back into university.

Daisuke: Actually, I started my original company Preferred Infrastructure. I did research at school at daytime and I did business maybe at night or at the morning, so I did not spend –

Tim: Okay. Well, that's right because Preferred Networks was spun out of Preferred Infrastructure. So you started Preferred Infrastructure while you were in college. Why did you decide to spin out a new company out of Preferred Infrastructure? Because it focuses on the same kind of technologies, right?

Daisuke: Yes. Preferred researcher focus is not using AI in the real world and we have very good business and the business was growing, so it was difficult to focus on two things, the current business and a very different business on AI and IOT. Therefore, we decided that it is better to separate the company into groups so that each group can focus on the one thing.

Tim: Okay, so if I understand, Preferred Infrastructure is more general AI and Preferred Networks is more applications and IOT?

Daisuke: Yeah. Preferred Infrastructure, the main business is to sell search engines or recommendation engine, and the main customer, the media companies and those main data was text.

Tim: Okay, so Preferred Networks is the one that has a more general mission to bring AI to different sorts of industries?

Daisuke: Yes, and the difference is Preferred Infrastructure, the original company, did not raise the money from the partners or venture capitals at all.

Tim: Okay, let's talk applications and some of the things you're working on. So one of the most exciting projects is your collaboration with Toyota and their focus on autonomous vehicles, and you've been working with them for three, four years now and they just recently invested $95 million into Preferred Networks to accelerate that research. So tell me about what you're doing with Toyota and the self-driving Prius.

Daisuke: We try to apply the AI technologies, especially the deep learning and the related technologies, and there are so many programs in the autonomous driving as you know. In our company,