Artificial intelligence was the study is 10 times faster and more efficiently

A division of Google, is engaged in development of artificial intelligence, announced the creation of a new method of training neural networks, combining the use of advanced algorithms and old video games. In the old Atari video games are used as a learning environment.

Artificial intelligence was the study is 10 times faster and more efficiently

DeepMind developers (remember that these people have created a neural network AlphaGo, repeatedly beat the best players in the logic of the game) say that the machines are able to be trained in the same way as humans. With DMLab-training system 30, created on the basis of the shooter and the Quake III Atari arcade games (using 57 different games), the engineers have developed a new machine learning algorithm IMPALA (Importance Weighted Actor-Learner Architectures). It allows the individual parts to study the implementation of several tasks, and then share their knowledge with each other.

Artificial intelligence was the study is 10 times faster and more efficiently

In many respects, the new system was based on an earlier architectural system A3C (Asynchronous Actor-Critic Agents), in which individual agents are investigating Wednesday, then the process is stopped, and they share their knowledge with the central component of the "pupil". As for IMPALA, then it may be more agents, and the learning process itself takes place somewhat differently. In her agent sends information to two "disciples", who then also communicate with each other. Furthermore, if A3C calculating the gradient of the loss function (in other words, a mismatch of the predicted and obtained parameter values) are engaged agents themselves which send information to the central core, the object of this system IMPALA engaged "pupils". An example of the passing game man:

Artificial intelligence was the study is 10 times faster and more efficiently

Here's how to handle the same task IMPALA system:

Artificial intelligence was the study is 10 times faster and more efficiently

One of the main problems in the development of AI is the time and the need for high computing power. Even in conditions of autonomous machines need rules by which they could follow in the course of your own experiments and finding ways of solving problems. Since we can not just build robots and make them available at will to learn, developers are using simulation methods and deep learning.

To modern neural network could learn something, they have to handle a huge amount of information in this case is - billions of frames. And the more they do it faster, so less time is spent on training.

According to the DeepMind, in the presence of a sufficient number of IMPALA processor achieves performance of 250 000 frames / s, or 21 billion frames per day. This is an absolute record for this kind of task, the portal The Next Web. Themselves as developers commented that their AI system to cope with the task better than comparable machines and people.

In the future, these AI algorithms can be used in robotics. By optimizing systems machine learning robots will quickly adapt to the environment and to work more efficiently.