On August 17, the world learned that Google had bought the Belarusian startup AIMatter, which had developed an image-processing technology based on neural networks. “We are excited to welcome the AIMatter team to Google,” said a Google spokesperson. What was it about the Belarusian Company that caught the attention of the IT giant?
AIMatter, founded by Yury Melnichek and Yuri Gurski less than a year ago, has released two apps – Fabby and Fabby Look. Fabby can separate a face from the background in a selfie – it can blur the background or replace it with cosmic landscapes, tropical wonders or kittens. Fabby Look can change one’s hair color. Both apps process photos and videos in real time, regardless of whether you are connected to the Internet.
Did Google really buy the apps just to use them for selfies? Not really. The apps have become a showcase for unique products by the company, which is a resident of the Minsk High Technologies Park. The main thing here is the possibility of quickly processing an image with the help of neural networks. According to the developers, they were the first in the world to implement this approach.
«There are even more important potential fields in which to use them – for example, in video messaging apps like Skype or in professional video editing apps,» says AIMatter co-founder Yuri Gurski. The company also created SDK – a set of development tools that allows the use of neural network technologies in other products.
AIMatter technologies can be used in many fields – for example, they can be used to remove noise from photos. Image: aimatter.com
What are neural networks? They are one of the directions in the development of artificial intelligence systems. Such systems are able to learn and to correct their own mistakes. They are called neural because the principle behind their operation resembles the human nervous system. In brief, a neural network consists of many subprograms – neurons, combined into layers. For example, one neuron can find lines, another one can find corners, and yet another can find color combinations.
Input data pass through the layers sequentially. Each neuron works like a light bulb: it comes on under certain conditions but does not under others. All neurons are closely linked, and their actions can change under the influence of tasks they perform – this is how the learning happens.
What are neurons able to do? They are able to do a lot. Imagine that you want to teach the network to recognize cats in a photo. For this you need to train it – you would take a large image database and pass it through the program, each time marking the correct answer.
Initially, neurons will work randomly, and the neural network will be able to tell a cat from a dog maybe half the time. However, the system will gradually train itself and single out the characteristics that define “cat.” It will make its own decision about which characteristics are determinative.
«We never told the network, ‘this is what a cat looks like’. By and large, it invented its own concept of cat,» said Jeff Dean of Google, who was involved in developing a neural network to search for cats in a YouTube video. As a result, the network will be able to recognize and categorize even images it had not seen before – almost in the same way that a human does.
Google’s AlphaGo program operates on the basis of trained neural networks. Recently, the program defeated a human rival in the game Go. Neural networks also underlie MSQRD and Prisma, two photo-processing programs that have made quite a splash, and which, incidentally, were also developed by Belarusian programmers.
Neural networks also control robots, compose music and draw pictures that many art experts are not able to tell from real ones. In addition, neural networks have already learned how to write their own computer codes. Some experts predict that neural networks will replace people in professions like translators, call center operators, security guards, and drivers.
What is so special about AIMatter neural networks that has so captivated the Americans? It’s the fact that they are quick on smartphones. Neural networks normally require lots of calculations. For example, there are 176 graphics cards per one AlphaGo program, and the neural network created by Google to search for cats involved 16,000 computers!
In AIMatter apps, all calculations happen on the same device. More precisely, training takes place on the developer’s servers, but the actual processing of photos or videos takes place right on your smartphone. Even though this neural network has a limited capacity and is not that much «smarter» than an ant’s brain, its potential uses can be very broad. Depending on your gadget’s model, apps developed in Belarus use the smartphone’s CPU (central processing unit) or GPU (graphics processing unit). The developers say that their technology is much faster than any other similar technology in the world.
«One neural network is able to mark and outline images of people in photos. Another several neural networks are trained for every style available in the app. To smoothly fit the silhouette of a person into some backgrounds, additional differential equations need to be solved – these calculations also occur on the phone itself,» Fabby ‘s developers explain.
This could be the reason that the famous IT giant bought the Belarusian company. «Rapid image processing is the only area where those giants lag behind small startups,» said Grigory Bakunov, director of technology distribution at the search company Yandex, commenting on the deal.
In line with the standard practice exercised by Google and many other corporations that buy startups, the amount of the transaction and other sensitive details will not be disclosed. «I can only say that we are very satisfied with the terms and conditions of the transaction and the way it went,» said project CEO Andrei Kulik.
«That was definitely not an acquisition. In other words, it was not a purchase of technologies, but a purchase of qualified specialists,» Yury Melnichek says. «We prefer the verb ‘join:’ AIMatter joins Google, which will allow much more to be done in the field of neural networks on mobile phones.»
«It was certainly not an acquisition,» agrees Yuri Gurski. «However, the people on the team are an important factor in technology absorption. Even at the risk of exaggerating or sounding ostentatious, I would say that only stars work at AIMatter and this is the strongest team in the field of artificial intelligence in Belarus. Our team has winners of international programming Olympiads, holders of international patents and authors of major works in scientific research. Many had worked for the top technology giants, like Google, Yandex, AMD, and NVIDIA. This is actually the first time that a major IT company like Google has bought a company from Belarus, and not an intellectual property or an affiliated company in another jurisdiction.
«A precedent has been set, and I think this is by no means the last such deal. There are an awful lot of talented engineers in Belarus. According to TopCoder statistics, our country holds 7th place in the global ranking of competitive programming, between the US and South Korea.»
How can you create a start-up that Google will buy? Many would like the answer to this question in the wake of the sensational news.
Yury Melnichek smiles:
«The main thing is to attract talented people to the company and not get in the way of their work.»
Credits: techcrunch.com, 42.tut.by, news.tut.by