Scientists decided to train artificial intelligence through games
AI has already made significant progress in analyzing information. However, technology is not yet capable, like a person, of understanding the context, the general message of the processed information. The researchers decided to teach artificial intelligence to this using games, where they should “draw” the phrase so that it could “decipher” the computer.
In the technology review, scientists from MIT draw public attention to the fact that artificial intelligence is already capable of beating champions in chess, go and many other games in Key StarCraft (more recently), but this remains a narrow area where AI is trained to win . In other words, if a robot consistently wins chess, it does not make progress in solving a sudoku. Scientists from the Allen Institute of Artificial Intelligence have developed a special service called Iconary, where anyone can relax the limitations that hold down computer thinking now. It is possible to play on the site iconary.allenai.org.
Do you want to train artificial intelligence by completing the game?
The whole point is that the person-player receives any phrase or description of the action, and this needs to be portrayed. Artificial intelligence analyzes the finished drawing and must guess. The more people play this game, the more the AI learns to understand the general meaning of the artistic context.
Frankly speaking, now participants need to succumb strongly to the computer. For example, AI will not be able to guess correctly a word or an action, unless each part of the phrase is turned into emoji. It would seem that if you make out the phrase “a man catches fish”, then it is enough to draw a person with a fishing rod and fish on it. However, today, the robot is not able to solve this. We will have to portray it in detail: the person, the fishing rod and the fish separately, then select the appropriate emoji from the menu and turn the finished phrase into images.
The developers have already integrated into the game there are 1200 smiles, about 75 thousand suggested phrases and a dictionary for 20 thousand words. In addition, you can choose between two modes of difficulty. The creators are planning to launch a table of the best participants. Thus, they are hoping for an influx of users who will compete with each other, and in the meantime, AI will learn to understand the artistic meaning of phrases better and faster.
One of the developers of this service, Ali Fahadi, is convinced that such games are much more useful and productive for the development of AI than board games, since in this case we need a much more developed AI. In addition, the game can become a kind of Turing test – Iconary participants can only guess with whom they are playing on the other side of the screen – with a person or a computer.
Note that in the summer of 2017, AI AlphaGo, developed within the walls of Google, defeated the strongest professional in the game Go Ke Jie. Remarkably, the game usually had a significant advantage, and Jie largely copied the strategy of the computer program itself.