Google AI creates its own ‘child’ AI that’s more advanced than systems built by humans

The AutoML project is said to make process of designing machine learning models 'much more accessible'

People are silhouetted as they pose with laptops in front of a screen projected with a Google logo, in this picture illustration taken in Zenica October 29, 2014

(REUTERS/Dado Ruvic)

Google has developed an artificial intelligence (AI) system that has created its own “child”.

What’s more, the original AI has trained its creation to such a high level that it outperforms every other human-built AI system like it.

It’s an impressive achievement, but one that could also trigger fears about what else AI could create without human involvement.

Google unveiled its AutoML project in May, with the aim of making it easier to design machine learning models by automating the process.

“In our approach..., a controller neural net can propose a ‘child’ model architecture, which can then be trained and evaluated for quality on a particular task,” the company said at the time.

“That feedback is then used to inform the controller how to improve its proposals for the next round. We repeat this process thousands of times — generating new architectures, testing them, and giving that feedback to the controller to learn from.”

In November, the AutoML plans were used to create NASNet, a “child” AI designed for object detection, which outperformed state-of-the-art machine-learning architectures built for academic competitions by humans.

To test NASNet, Google applied it to the ImageNet image classification and COCO object detection dataset, which it describes as “two of the most respected large scale academic datasets in computer vision”.

On ImageNet, NASNet achieved a prediction accuracy of 82.7 per cent, performing 1.2 per cent better than all previous published results.

On COCO, Google says NASNet achieved “43.1% mAP which is 4% better than the previous, published state-of-the-art [predictive performance on the object detection task]”.

“We hope that the larger machine learning community will be able to build on these models to address multitudes of computer vision problems we have not yet imagined,” said the researchers, who have open-sourced NASNet so it can be used for computer vision applications.