Google AI engine can replicate itself without human help

Google has unveiled an AI tool that can replicate itself and compete with the human child

American tech giant Google's AutoML artificial intelligence engine has long been capable of creating AI solutions without the support of humans - and it's just completed its biggest challenge to date.

After a team of Google researchers used reinforcement learning to automate machine learning models, the AI was able to build a virtual child more powerful than human youngsters.

Unveiled in May 2017, AutoML is an artificial intelligence tool that can learn to replicate itself and build new AI solutions. In the case of the child AI, it acts as a so-called controller neural network.

The bot, called NASNet, is capable of identifying a variety of objects presented in a real-time video. For instance, it can recognise traffic lights, backpacks, handbags, people, cars and traffic lights.

Google explained in a blog: "A few months ago, we introduced our AutoML project, an approach that automates the design of machine learning models.

"While we found that AutoML can design small neural networks that perform on par with neural networks designed by human experts, these results were constrained to small academic datasets like CIFAR-10, and Penn Treebank."

Although NASNet is far from perfect, AutoML has the ability to make constant improvements. It monitors the child bot's performance and looks for areas where it can improve.

Google researchers put the child bot up against the ImageNet image classification and COCO object detection datasets, which are powerful academic data resources, to test its computer vision abilities.

And the results showed that NASNet can outperform the majority of existing computer vision systems, showing an 82.7 per cent accuracy in predicting images.

Machine learning is a lucrative area, but it does take up a lot of skill and time. Google believes that AutoML could bring the concept to people who aren't necessarily AI and ML experts.

"We suspect that the image features learned by NASNet on ImageNet and COCO may be reused for many computer vision applications," added the researchers.

"Thus, we have open-sourced NASNet for inference on image classification and for object detection in the Slim and Object Detection TensorFlow repositories.

"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."