What is more, the original AI has trained its own creation to such a high degree that it outperforms every other human-built AI system like it.
It is an impressive accomplishment, but one that could also trigger fears about what else AI can create without human involvement.
Google unveiled its AutoML project in May, with the aim of making it easier to style machine learning models by automating the procedure.
“In our approach…, a control neural net can propose a ‘kid’ model structure, which may then be trained and evaluated for quality on a certain task,” the company said in the time.
“That feedback is then utilized to inform the controller how to boost its suggestions for the next round. We repeat this process thousands of times — generating new architectures, testing them , and giving that feedback to the controller to find out from.”
Back in November, the AutoML programs were used to make NASNet, a “child” AI constructed for object discovery, which outperformed state-of-the-art machine-learning architectures built for academic contests by humans.
To examine NASNet, Google applied it to the ImageNet picture classification and COCO object detection dataset, which it describes as “two of the most respected large scale academic datasets in computer vision”.
About ImageNet, NASNet achieved a prediction accuracy of 82.7 per cent, performing 1.2 per cent better than most previous published results.
On COCO, Google says NASNet attained “43.1% mAP that’s 4 percent better than the previous, published state-of-the-art [predictive performance on the object detection task]”.
“We expect that the larger machine learning system will be able to build on those models to tackle multitudes of computer vision problems we’ve not yet imagined,” said the investigators, that have open-sourced NASNet therefore it can be used for computer vision software.