Now everyone can become a machine learning professional.

Mark Hammond at Bonsai headquarters in the suburb of Berkeley

Do you successfully play football, shoot in a popular movie, or successfully play the stock market? Congratulations - you are almost as valuable as a data processing or machine learning specialist with a doctoral degree from Stanford, MIT or Carnegie Mellon. At least it looks that way. All companies in Silicon Valley - and in principle, already and not only there - are feverishly competing to get such a man-prize. This is something like truffle hunting performed by HR managers. As enterprises realize that their rivals rely on artificial intelligence (AI) and machine learning (MO), the number of vacancies for these specialists will constantly increase.

But what if you could get the benefits of AI without hiring expensive and talented specialists? What if smart software drops a plan? Is it possible to get the benefits of deep learning (GO) without special talents?

Startup Bonsai [Bonsai] is one of the companies that respond positively to this question. With their help, perhaps, the democratization of AI is coming, which as a result may affect millions, if not billions, of people.
')
Today at O'Reilly's AI conference in New York, Gen. Bonsai, Mark Hammond will show a demonstration of how the system works. The demo reproduces one of the iconic achievements of HE: how the DeepMind system learned how to play old Atari toys. Specifically, the Bonsai program will learn to play Breakout, where the platform beats a ball that destroys a wall of bricks (the game from 1976 was a breakthrough at one time - by the way, Steve Jobs was working on it).


37 lines of code create a neural network learning to play on their own

At the same time, DeepMind created world-class AI geniuses and trained the neural network on a set of games from Atari. Such an achievement has been credited with publishing in a world-class magazine. In the case of Bonsai, this is more a way to cut a path. It starts with a development system located in the cloud. One programmer, perhaps even having never attended an AI course in college or on the Internet, can describe the game, and the system will choose the most optimal algorithm for learning the neural network. Then the programmer in a few minutes encodes the concept of the game - for example, the need to keep the platform under the ball - and gives Bonsai the opportunity to work independently with neural networks, optimizing them for the best result.

The version of the game from Bonsai takes 37 lines of code. But this is deceptive. When Hammond shows what is going on “under the hood” of the system, he demonstrates a graph showing how the system builds a complex neural network worthy of one of the ninjas who are involved in DoD at Google. The programmer will not have to deal with all this MO-good.


The trick is amazing. “I’m usually not very impressed with demos,” said George Williams, a researcher at the Courant Institute of Mathematical Sciences at New York University. “But what Mark showed me was both plausible and simply amazing.” He understood our current state with MO and the tools needed to create the next generation AI. ”

Whether Bonsai will become the leader of this trend is still unknown. But Williams is right about “where we are” on the scale of AI development. The next step is the emergence and flourishing of smart computers with MO for dummies.

Bonsai was born on the beach. Hammond, who previously worked at Microsoft as a developer, has been interested in AI for some time. After leaving the company in 2004, he worked at Yale University in the field of neurology. In 2010, he spent some time at Numenta's AI startup under the direction of Jeff Hawkins (co-founder of Palm handheld computer), but left to launch a third-party company, which he subsequently sold.

Then in 2012, Hammond was visiting friends in southern California. His toddler son was tired, and the whole company was returning to the car. While Hammond's wife chatted with friends, and the son fell asleep in her arms, he conducted a thought experiment. It began with the meme popular in the world of AI - the concept of the “main algorithm”. As suggested by Professor Pedro Domingos of the University of Washington (in the book of the same name ), this not yet open MO technology will be a universal solution for all problems. When scientists derive this algorithm, we can apply the AI ​​to anything.

But Hammond found a flaw in this reasoning. Suppose we find such an algorithm. Who will adapt it to countless practical applications? Today only adepts of the MoD are capable of this. There are too few of them, and too many tasks. We need a system that will lower the bar so that an ordinary developer can use these tools. Such a system does not require a narrow specialization in MO to train neural networks, but will allow programmers to learn the system to produce the desired results.

Gradually, he made an analogy with the history of programming. Initially, it was necessary to write programs in machine code. Then the programmers developed a standard instruction set, an assembler. A breakthrough came with the development of a compiler that translated high-level languages ​​into assembler. And after that, programming began to allow beginners to create serious programs. Hammond believes tools like Google's TensorFlow resemble an assembler era. They facilitate the construction of neural networks, but still the entrance to this area is accessible to those who are well versed in the work of neural networks. He wanted to create something like a compiler to expand this input.

He shared the idea with Keen Browne, a former colleague at Microsoft, who recently sold his gaming start to the Chinese. Tom liked the idea because he himself tried to comprehend deep learning using popular means available. “I'm pretty smart,” he says. - I was in China, learned their language. I programmed in Microsoft. But this occupation turned out to be ridiculous. ” And he subscribed to the Bonsai co-foundation. This name was chosen because Japanese plants, artificially bred, keep a balance between natural and artificial. As a bonus, managed to register the domain bons.ai.

Bonsai is not the only startup trying to solve the problem of lack of specialists in AI. Some companies started training their own employees on neural networks. Google has developed a whole range of internal courses, and Apple is looking for programmers with skills that would allow them to learn the right subject without problems. Google also released the TensorFlow program, which helps their engineers build neural networks. There are other tools for working with AI, and more likely to follow them, with varying degrees of understanding of the issue.

There are other startups involved in democratizing AI. Bottlenose appeals to a different audience than Bonsai: they work not for programmers, but for business analysts. But the promise is familiar. “We are giving new opportunities to users who are not scientists or programmers,” says the company's general director, Nova Spivack [Nova Spivack]. Other startups are taking even broader: at the conference, the head of the company Clarifai gave a talk “How to allow anyone to train and use AI”.

So, although Bonsai seems to have found a good niche in time, it will be difficult to draw attention to oneself due to the vigorous activity in this area. Adam Chayer, an AI expert, one of the creators of Siri, and the chief programmer at Viv, is impressed with the startup product. But he notes that although Bonsai makes AI closer to newbies, they still have to strain their brains and learn its programming language and system operation. “When big companies like Google roll out the system, people get off their feet to figure it out,” he says. - But a startup is not so easy to interest people. Do they have enough power to get enough users to become popular? ”

According to Hammond, building neural networks with the help of Bonsai is different from how professionals do it, at key points. Now, to solve a specific problem, it is necessary to select the right tools, which requires experience and knowledge. And Bonsai will have to do it himself. You only need to describe the concept of what you want to train the system.

And while experienced professionals will teach the network, comparing their output with the desired results, Bonsai will allow you to learn the system, breaking the process into concepts. For example, if you want the system to recognize a photo of dogs, you can characterize a dog as having four legs, a long muzzle, a long tongue hanging from the mouth. You are pushing the system, and the “intellectual center” in the cloud itself understands everything.

There are pluses. Scientists often do not understand how well-trained networks do their work, since they themselves set up and organize all concepts in an incomprehensible way. But in Bonsai, user-defined concepts give us a map of the thinking of a neural network. “The program should not be a black box,” says Hammond. For example, if you are programming a robot and the machine does not hit the brake in time, you need to be able to figure out what your system thought at that moment.

The question is whether such an abstraction will not lead to a drop in speed and efficiency. Usually with compilers it happens - programs do not work as fast as those written in assembler. Doubt is also caused by the ability of the system to choose the right tools for solving problems no worse than the doctors of science whom it must replace.

“I think that compromise cannot be avoided,” says Lila Tretikov, an AI specialist who previously worked at the Wikimedia Foundation and advised Bonsai. “This is not exactly the same thing as having a team of doctors of sciences on hand. But what is more important - uncompromising or just the opportunity to do what is necessary? ". Adam Chayer from Viv believes that Bonsai may not work as effectively as a system optimized for a specific task. “But its code is good, and it allows you to be at the top level of abstraction,” he says. Chayer says that his company has several valuable AI specialists, and therefore they are unlikely to use Bonsai - unless, as a tool for prototyping.

Bonsai tries his hand at tasks that have not yet been solved by classical systems with AI. “We are working on different games,” says Hammond, explaining that games are keys to several major issues that Bonsai is planning to approach. “Certain classes of AI games have not yet got to the core, even DeepMind. They trained on a bunch of games other than Breakout, but, for example, they didn’t manage to get the system to play Pac-Man successfully. ”

But the main thing is how Bonsai joins the movement to transfer AI into the hands of people who do not have much knowledge in this matter. It can be expected that many high-level tools will become more powerful and ubiquitous. Will we get to the point where every person on the planet trains and uses AI? At least, many clever people with money put on it.

“We have analysts in the cloud,” says Spivak, General Director of Bottlenose. He says that these virtual consultants can be called with questions like "which college should I go to." The cost of the system is nominal, and maybe zero. “It’s not possible to defend yourself in a bad decision by the fact that you cannot afford AI,” he says.

Maybe we will get even to the point where the AI ​​successfully conquers Pac-Man. Bonsai has not yet coped with this. “We are working on this,” Hammond says. “There are no announcements on this issue yet.”