China is overflowing with deep learning and AI startups, and they’re doing things differently from deep learning startups in Silicon Valley.

This video features Dr. Chris Rowen, founder and principal of Cognite Ventures and co-founder and CEO of Babblabs.


Dr. Chris Rowen: “I think that there a set of common features and a set of distinct characteristics about what the Chinese neural network startups are doing relative to the US, relative to the rest of the world. Some of the clear themes are, number one, there’s lots of interest in video and especially in surveillance in China.Some of that comes from the level of investment that the Chinese government is making in public safety. Some of it, I think, comes from a long history of interest in imaging.

“Some of it comes from the fact that the Chinese electronics community really understands a lot about devices. Gadgets. Things that are widely deployed in high volume.

“And there is, on the flip side, relatively less activity in the cloud in China among startups. Certainly it’s true in the neural network area. In fact, worldwide in my Cognite Ventures 300 list roughly two-thirds of all of the neural network startups are deploying software for the cloud. In China, very few.

“So it’s almost as if the whole cloud segment is missing from the Chinese neural network community. Partly because I think the cloud is less well-developed as a commercial ecosystem. Certainly there are big players like Alibaba, for example, and Baidu, focused on the cloud. But it’s really percolated much less completely into the ecosystem of small, medium, and large players.

“I think some of it really reflects the fact that people see the value in these things where there’s a tangible element. It may go back historically to the question of intellectual property protection in that it’s much easier to be confident that you’re protecting the value of your product when all of the software and all of the hardware are bundled together in one thing. And so that tends to make the Chinese startup community much more oriented towards hardware/software systems.

“It’s much less about ‘Oh, I have a very specific algorithm or a very specific application or a very specific chip in mind’ but rather: ‘I am combining chips and software and algorithms and application insights together in a package which is salable.’ So in an ironic sense, the Chinese deep learning startups tend to be quite integrated and quite system-oriented in their perspective. Because, I think, it reflects in part this interest in delivering a more complete solution, in part to protect intellectual property; certainly to protect value in a world where there [is] a long history of clone makers of various kinds.

“In general, the level of activity and the level of sophistication is high in China just as it is in the other leading countries that are at the forefront of this. I think some of the unique characteristics of China are the things I’ve mentioned; this focus on embedded devices, this focus on surveillance, the relative immaturity of the cloud. You also, I think, have an environment where there’s actually quite a bit of venture money available. The level of enthusiasm, the kind of overt encouragement of the Chinese government for investment in this really leading-edge technology is [making] a lot of money available to teams so that, while it’s hard to do an apples-to-apples comparison, my guess is that for a team of a given caliber they probably have the easiest time getting funded in China and get the highest valuations in China right now.

“I tend to be a big fan of the company DeePhi which is focused on really quite advanced neural network methods for vision, applying it, of course, in surveillance but in other areas as well. They are both developing new algorithms and optimizing them very effectively for FPGAs but also working to develop new silicon platforms.

“And they have, I think, a very significant history in working in a particularly interesting problem in neural networks which is how do you build much smaller, much leaner much more efficient neural networks than what people are typically implementing in the cloud. And they’re able to do things that have significantly lower compute requirements, fit into smaller chips, run at higher framerates and still keep very high accuracy. Accuracy comparable to what people have done in cloud-based computing, but now doing it in quite small devices.

“I think it’s a really interesting group of people out of a combination of Tsinghua University and Stanford who are the inspiration behind DeePhi and they’re doing quite remarkable things.”