Google’s AI has created a ‘child AI’ that can outperform itself


Artificial Intelligence (AI) created by humans is quite believable but an AI created by another AI seems like a thing straight out of a futuristic science fiction movie.

Researchers at Google Brain recently presented AutoML which is an automated AI that created an AI child called NASNet which could recall objects with 82.7 percent accuracy. AutoML teaches NASNet how to do specific tasks.

Basically, it is “an AI that can create an AI better than the AI it is.”

NASNet learns things through reinforcement learning, which means it reports back to its AI parent (AutoML) and learns how to do the task better.

As of now, it recognises objects, such as people, cars, handbags and traffic lights, in real time.

Google researchers mentioned in a blog post that they hope developers will be able to create “multitudes of computer vision problems we have not yet imagined”.

Image detection by the AI (Courtesy: Google Research Blog) 

AutoML was created using reinforcement learning, reports Futurism. It is able to detect objects and identify them with a significantly higher level of accuracy than a neural network designed by a human.

To design AutoML, the team used Imagenet and COCO- described by Google as “two of the most respected large-scale academic data sets in computer vision”.

Researchers have claimed that NASnet (the child Neural Network) was 82.7 percent accurate in predicting images from the validation set- which also happens to be a 1.2 percent jump from any previous attempts by a machine programmed AI.

The project is going on even as the discourses on whether Artificial Intelligence being safe for the world continues. The major Mark Zuckerberg Elon Musk spat on Twitter is a popular reminder of it.