Google makes super AI with imagination like people

DeepMind of Google has revealed a new research project to create imagination for artificial intelligence (AI).

This project wants a breakthrough that makes the artificial system think about their actions and make "cautious inferences". The advanced system using internal 'imaginary encoders' helps the AI ​​decide what useful and unproductive predictions about its surroundings.

"Imagine the consequences of your actions before implementing them is a powerful tool in human awareness ," DeepMind said.

Picture 1 of Google makes super AI with imagination like people
Who will be able to imagine, reason like humans?- (Photo: Corbis).

For example, when placing a cup next to the table, we will stop to see how stable it is and whether it will fall. Based on that imaginary result, we can re-adjust the glass to prevent it from falling and breaking.

According to DeepMind, this form of cautious reasoning is mostly " imaginative " , and is a distinct ability of people. This is also an important tool in our daily lives.

DeepMind revealed, the work was done in AlphaGo project. AlphaGo uses ' internal model' to analyze actions that will lead to future outcomes that can be deduced and planned.

DeepMind said: "These internal models work well because the environment like Go is 'perfect' - in which clearly defined rules allow the results to be accurately predicted in almost every situation.

But the real world is more complex, rules are not clearly defined and unpredictable. Even for the smartest machines, being able to imagine in these complex environments is a long and expensive process. '

Picture 2 of Google makes super AI with imagination like people
The process of forming imagination for AI takes time and money - (Photo: IT).

This is also the research team that revealed the work of teaching AI to walk. Researchers have trained a number of simulation bodies - including a headless walker, a four-legged ant and a 3D model - to learn complex behaviors when making them work. Different moving cases.

The results show how these systems can learn to improve their techniques when they interact in different environments so they can finally run, jump, and rotate.

In a new article published on arXiv, Google's DeepMind explains how simple functions can lead to rich and powerful behaviors. It is known that they have set up simulations for these characters to overcome some obstacles to help the AI ​​learn the best way to go from point to point.

Initially, when the characters run in each simulation environment, they are like drunken people, stumbling, colliding with walls . But over time, they become more intelligent, can navigate well in the Different types of terrain.

Thus, artificial intelligence is still improving and these capabilities will soon allow the machine to perform much more complex tasks than it is now.