Google Brain, an artificial intelligence and machine learning project at Google, has been used to power services like Android’s speech recognition system and photo search on Google+.
Now, two of the most longstanding machine learning engineers, one of whom worked on Google Brain, have left the search giant to start a new company. The idea: to build machine learning, artificial intelligence technology similar to what’s used internally by companies like Google, making it available as a cloud service that can be used by anyone.
An AngelList profile lists the two founders, Olcan Sercinoglu (first started to work for Google in 2001) and Dmitry Lepikhin (an eight-year Google vet who worked on projects like Sibyl, widely used in products like YouTube and Google Play). It also notes notes that the startup has already raised $4 million in seed funding from investors that include distinguished Google engineers Georges Harik and Gautham Thambidorai, Felicis Ventures, Amplify Partners, Tencent’s David Wallerstein, Tencent itself, and others.
Contacted about the startup, Sercinoglu tells us that Quora’s Adam D’Angelo is also among its backers, that the seed round is likely to be around $8 million when it closes, and that they are now on the hunt for engineers and others to flesh out the business.
There have been a number of notable departures from Google Brain — a so-called Google Brain Drain as one person who I spoke to described it. Andrew Ng, one of the original founders and leaders of the project, left to co-found Coursera and earlier this year joined Baidu. Quoc V. Le is thought to have left as well.
Other Googlers with deep learning, object recognition and image classification expertise have left for Facebook, to fill the ranks for the social network’s own artificial intelligence project. And when you consider that companies like Apple are also ramping up in search, there is clear demand in the market for deep learning expertise.
For Sercinoglu, this is actually the second time that he has left Google — the first being when he worked on another startup called Imanen, focused on intelligent mobile applications that adapted based on user interactions. Sercinoglu says that Imanen’s IP is now being transferred to Scaled Inference.
Sercinoglu and Lepikhin chose to build this project as a startup rather than within Google, even though they could have stayed there to work on it.
“We had the opportunity to do this exact project at Google,” he says. “They were willing to acquire the IP from Imanen to let me work on this there. But I just felt that it might be a great opportunity to build a company like this, to sustain itself.”
It’s very much a case of no hard feelings. Sercinoglu has a history of doing his own thing: while still at university he also built a music search engine called AVFind.com, which is no longer running but had been the first project of his that helped catch Google’s attention.
“I’ve always had a tendency to build my own thing, and how long I stayed at Google speaks to how great Google is.”
So what is Scaled Inference doing?
While startups like DeepMind (acquired by Google) and Vicarious (still independent with significant backing) are focused on the long term problems of human-level AI, imitating the human brain with machines, Scaled Inference is based around the idea of takings this concept and effectively making it a neutral platform for different kinds of problems.
(Here is one analogy: if you think of individual AI applications as vertical, specific search engines, then Scaled Inference positions itself as the Google in the world of artificial intelligence.)
Scaled Inference positions itself differently from these other companies, too. “We believe that an AI company should be able to provide useful services to sustain itself along the (very long) path to advanced general AI,” Sercinoglu says, “and also that ultimately such AI does not need to imitate humans any more than an airplane needs to flap its wings.”
Scaled Inference will provide its AI as a cloud service — virtually any kind of machine learning and AI that can be used on the web or in apps, particularly those that are built around creating services that are based on individual users and a particular context, be it location, browsing history or something else.
Mobile apps, created for smartphones with small screens and people on the move, are particularly suited to this idea of doing the legwork for users. It’s something that even companies like LinkedIn are trying to tap into more, specifically for consumer-facing services.
“Big companies have been using machine learning and AI for a long time and it’s not new but it’s becoming more visible now,” Sercinoglu notes. “We would like to power the new generation of apps that are intelligent and more context-sensitive.”
There will be other areas where Scaled Inference’s platform could apply. For example, an application Sercinoglu released with Imanen called River aggregated a user’s various feeds from Facebook, Twitter, and so on, and then learned to rank items according to the probabilty of a user interacting with them — not unlike how Facebook moderates its own feed these days.
“River would capture your intent and reduce features that can be fed into the AI system to get a ranking of items that you haven’t read according to the probably of you engaging with them,” Sercinoglu says. While River was built to run on devices the idea is to take some of that IP and put it into the cloud.
When launched, Scaled Inference will be based around a set of APIs, three of which will be released at launch.
The first set, the “lowest level” as Sercinoglu describes it, will be around pattern reognition. “You submit data to us and we start recognising patterns and start alerting you to patterns or anomalies in it,” he says.
Another will let developers predict a user’s actions based on questions that are answered. (Something that you could imagine might apply in customer service surveys, for example.)
And the third will be a ranking product based around the basic functions that existed in River. (Think here of how such a service could apply to a site like Quora, for example, where people search for information and this could help rank the results.)
Quora has its own in-house solution for ranking search results, but with D’Angelo investing in Scaled Inference, the company is also looking at what Scaled Inference is creating.
The whole platform, in fact, will be built with some of its own AI in mind: “We will offer these three sets of APIs and our goal is to make them simple and intuitive,” he says. “You will only need to describe the problem — no mention of the methods that you should use. Our systems should be able to figure that out based on what you have.”