Following "AlphaGo" Google 's DeepMind is developing a "superhuman AI" with human intelligence


Google 's Artificial Intelligence Development Company "DeepMind" that developed Go AI that has made the world' s strongest players develops AI with human - like recognition capability that predicts the behavior of objects such as "dogs chase cats" It is.

DeepMind's neural network teaches AI to reason about the world | New Scientist

Forget AlphaGo-DeepMind Has a More Interesting Step Toward General AI - MIT Technology Review

When choosing the best banana in the supermarket, humans look at the multiple bananas lined up on the shelf and process various information such as "the right banana is colorful", "the left banana is bigger than others" I will. The ability to infer the causal relationship between such objects, words, and ideas is called "relational reasoning". Although the artificial intelligence camera system can distinguish between dogs and cats, it can not have an intuitive perception that "a dog chases a running cat", and relation inference is considered difficult for AI.

DeepMind's research team has developed two systems with human-like recognition capabilities such as relationship inference and publishes their respective papers. The first system uses the data set "CLEVR" of the object, and the research team has the system train object recognition by machine learning and then "whether one object is in front of another object" I asked for a positional relationship such as "which object is closest". The system succeeded in leaving dramatically better results than any system developed so far, in some cases it exceeded human performance.

The other system is a similarly trained machine learning system that predicts simple object behavior in two dimensions. This is more sophisticated than "inferring objects in one scene" as used in the first paper, but human beings predict and catch behaviors of flying balls in three dimensions It is based on these forecasts that we are able to drive a car because we are doing complex processing such as. In one psychology research, we refer to human inference ability as "intuitive physics engine".


Although these DeepMind machine learning systems show that it can demonstrate human-like recognition ability in local tasks, it is not yet a striking breakthrough. Professor Sam Gershman, who is studying intelligence at Harvard University's psychology department, says, "Superhuman performance in certain machine learning tasks does not indicate" superhuman intelligence ". However, there is no doubt that it is a research in fields required for the future development of artificial intelligence, it is expected to become a clue to realizing "superhumanistic AI" as seen in SF movies.

in Software, Posted by darkhorse_log