Many of us remember and know what kindergartens are like. The training rooms are stuffed with toys and puzzles, music and books, flowers and even sometimes cats, thus revealing a rich and lush world with which children can play and learn while playing. But contrary to popular belief, the game of children is far from simple. They are not just having fun - they are learning, gaining an understanding of the world. Due to the game in a diverse and delightful universe, we cherish the child's polygonal intelligence.
Why not teach AI in the same way?
Not so long ago, the non-profit institution OpenAI opened a virtual world for AI, so that it could explore and play with it. The goal of the Universe project is as big as its name: to teach a single AI to perform any task that a person can perform using a computer.
Teaching individual AI agents to succeed in various real-world tasks, OpenAI hopes to bring us a step closer to truly intelligent bots - who have flexible reasoning skills like the ones we have.
Intellect of the general level
There is no doubt that AI becomes frighteningly intelligent.
Now computers can accurately see, hear and translate languages, sometimes even ahead of people. Only at the beginning of this year, in a series of high-profile games in go, AlphaGo from DeepMind crushed 18-times world champion Lee Sedol, ten years earlier than some experts had expected.
But the truth is that AIs are just as good as they were taught. Ask AlphaGo to play chess, and the program is likely to get lost in the machine, even if you explain the rules to it in the smallest detail.
So far, our AI systems are ultra-efficient horses for one trick. The method of teaching is partly responsible for this: the researchers begin with a clean sheet of AI, conducting it through millions of tests, until he succeeds in one task and does not solve it. AI never experiences anything else, since he learns how to solve any other problem?
To get to the intelligence of the general level - able to use the experience at the human level to solve new problems - AI needs to transfer their experience to other tasks. And that's where Universe will help them. Experiencing a world full of different scenarios, the OpenAI scientists hope that AI will gain knowledge about the world and flexible problem-solving skills that will allow it to "think" rather than stuck forever in a single loop.
Brave New World
At its core, Universe is a powerful platform that includes thousands of environments, usually providing standard methods for training AI agents. As a software platform, Universe provides a platform for running other people's software, so that programs are trained in different environments - Atari and flash games, applications and websites, for example, are already taken.
There will be others ahead.
In theory, Universe can run any software on any computer, allowing scientists to insert and train their AIs at will. It's like sending a child to a summer camp: you choose your niche, type of activity, wait until he learns it, then another, and so on, swim and repeat.
In the Universe, AI interacts with the virtual world in the same way that people use the computer: it "sees" the pixels on the screen and uses the virtual keyboard and mouse to enter commands.
This is made possible by Virtual Network Computing (VNC), in fact, a desktop sharing system that allows you to transfer keyboard and mouse movements from one computer (AI) to another (learning environment). When the environment changes, the VNC sends updated screenshots back to the AI, allowing it to complete the next step. VNC acts like the eyes and hands of AI.
How does the training work?
All AIs that are connected to the Universe are trained through so-called reinforcement training, a powerful method that led to the success of AlphaGo. This term hides, however, the way we humans train dolphins, dogs and even children. This training by trial and error: select the action, and if you are rewarded for it, continue in the same spirit. If not, try something else.
Instead of starting with a completely empty AI, researchers sometimes give them an impulse, allowing them to "look" as people solve the problem. This allows the AI to form a first impression and have a more complete idea of how to optimize their decisions.
Training with reinforcement is already used in many AI applications. Inside Universe, however, the power of this technology is revealed to the fullest. Because AI can jump between games and applications, it can take what it has learned in one application and easily use it to understand another - it's called "transfer training" or "transfer training". This skill is not easy to master, but it is necessary on the road to intelligent machines.
According to OpenAI, we are slowly moving there: some of their agents are already showing signs of transferring learning from one game to another.
From games to the world of bits
Like many other AI developers, OpenAI uses games to push Universe, for a reason: it's easy to evaluate them from a position of success. Since games are measured by different statistics and estimates, the system can easily use these figures to evaluate the progress of AI and reward it accordingly. This is extremely important for training with reinforcements.
Because Universe relies on pixels and keyboards, people can also play games on the platform. These sessions are recorded and provide a basic level for assessing the performances of AI (good work, you will agree).
But games are just a small part of our interaction with the digital world, and Universe already goes beyond its limitations with the Mini World of Bits project. Bits are a collection of various interactions with browsers that we encounter, a furrow of the Internet's abyss: when you enter text or select options from the drop-down menus, by pressing "send".
These tasks, though simple, form the foundation for how we connect to the treasury called the Network. OpenAI wants AI to move freely on the Internet - for example, it could book a plane ticket. In one of the Universe environments, researchers already give the AI the desired booking schedule and teach it to search for flights on various airlines.
And this is only the beginning.
Universe is only growing and expanding. Microsoft's Malmo platform, which uses Minecraft for AI training, must integrate with Universe. A popular game with folding protein fold.it, Android applications, HTML5 games and more are waiting in line.
Ghost in the car
So, now we can teach AI to play various games and browse the web. Great business, you think. Will this lead us to a general level of intelligence?
Perhaps, and the road will be long.
But AI, who knows how to win in any game that you throw to him, can only think logically and in several steps to achieve victory. AI, which can move around the chaotic world of GTA V, must already understand the basics of real-world physics, brutality and retaliatory measures. AI, which can work on the Internet, already knows how people usually communicate with each other and can use this knowledge to get information, create their own web identity or even look into yours.
Every day we learn, play, work and grow in the digital realm. For many, the world of zeros and ones is just as real as the one in which we were born. Now that AI has access to this digital world, it's its turn to grow. Let's see how far he can go.
The article is based on materials .
Comments
Post a Comment