Announcement that Artificial Intelligence of Google "Deep Mind" learned "ability to touch and remember" like a baby

ByWarren R. M. Stuart

In order to know the properties such as the weight and fragility of the object in front of you, if you are a human, you first understand what you are doing through the action of "touching" and "grabbing" Especially when you are a curious child, you may take action like "try to put it in your mouth". Through such behavior like "human beings" trying to grasp the characteristics of objects in front of youGoogle Deep MindIt is announced that the artificial intelligence (AI) of the success was successful.

(PDF)1611.01843 v1.pdf


Google DeepMind's AI learns to play with physical objects | New Scientist
https://www.newscientist.com/article/2112455-google-deepminds-ai-learns-to-play-with-physical-objects/?

Google teaches DeepMind AI how to interact with real objects; squish puny humans
https://www.neowin.net/news/google-teaches-deepmind-ai-how-to-interact-with-real-objects-squish-puny-humans

A research team by Google and the University of California at Berkeley will be announced in the future, according to a peer-reviewed article, Google DeepMind has succeeded in developing an AI robot capable of grasping the weight of an existing object about. In other words, this means that the robot can now take actions that humans take to grasp the nature of the object, and if the robot exceeds the human ability "SingularityIt is an event that it can be said that I climbed the step for reaching (technical singular point) again.

In the research, five blocks of the same size and different weights are placed in a row in front of the AI ​​robot, and learning is repeated to let the robot apply the heaviest one to the robot. If you choose the right one, you will be given programmatic rewards, and if you choose the wrong one, negative feedback will be given and the next action will be improved. Based on these, the AI ​​robot will try various ways that you can get a good reward.

Initially it is an AI robot that he chose a block as "appropriate", but while repeating learning, it is the only way that the only thing that comes to the correct answer is to "Choose all the blocks and pick the heaviest thing" He said that he learned something.


Even for tasks given to the robot, up to five blocks are used to select the heaviest one as well, but this time the block arrangement has been changed to stack in tower rather than one row I will. First of all, the robot tried to grasp how many blocks are there, and eventually he took actions to grab all the blocks and find the correct answer in the same way. Among them, the robot seems to have learned the act "to break the stacked blocks by pulling the tower".


As AI robots do machine learning based on rewards and negative feedback like this timeReinforcement learningIt was also used for learning of Deep Mind's "AI that gained better than humans" released in 2014.

When will the day the artificial intelligence "DQN" that learns the game by himself learns more than humans threaten humans? - GIGAZINE


By carrying out reinforcement learning, AI will be able to find a solution without human's or animal's special instructions. Professor Eleni Vasilaki of the University of Sheffield said, "With this ability, we will be able to find solutions in new and clever ways to deal with known problems and even without human instruction." I will.

This ability is also a small step in technological progress toward singularity, but what will become one big foothold in the future when robots and AI act in the real world It is said that it can be said that it is.