Google is betting big on artificial intelligence (AI), and it’s clearly paying off. Apart from offering up collections of code that best the world’s board game champions, they’ve also managed to create an AI that, in effect, designs its own AI – and its creations have gone from analyzing words to disseminating complex imagery in a matter of months.
On a company blog post from May of this year, engineers explain how their AutoML system (Automated Machine Learning) gets a controller AI – which we can perhaps call the “parent” in a colloquial sense – that proposes designs for what the team call a “child” AI architecture.
The child is then given a task, and feedback is sent to the parent. This allows the parent to improve how it designs a second child, and so on and so forth, thousands of times over. This self-reinforcing learning mechanism allows it to develop AI children that ultimately are better than anything human engineers can make.
“Our approach can design models that achieve accuracies on par with state-of-art models designed by machine learning experts (including some on our own team!),” the post notes. They even add that the AI children have certain design features that seem to be of no clear use to their own researchers.
The AutoML system has cropped up in the news again, and although much of the focus has been on the ability to create impressive AI children, a recent blog post by Google reveals that there’s more to it than that.
The data sets that AutoML focused on were relatively simple, and over the last few months, engineers have been looking at how it deals with image processing and recognition. If you’ve ever had to log in to a website recently and had to prove you’re not a robot, chances are you had to click on a series of images to demonstrate this.
At present, we’re better than AI at dealing with images. Our days may be numbered though, because AutoML can now pick out specific objects in images better than any other computer vision system.
For example, say you’ve got an image of a person climbing a mountain. AutoML’s new “child” AI, dubbed NASNet, can – to an 82.7 percent accuracy rate – pick out the individual elements; a person, their walking stick, their backpack, their clouds, the Sun, and so on.
A different child AI did this a few months back, using data sets that featured catalogues of words and color images. The new data sets, however, contain far more detailed and numerous images, and NASNet is nailing it.
Remember, this is an AI that another AI effectively created itself. That’s an incredibly exciting advance.
Until recently, when it comes to AI, much of the media attention has focused on DeepMind’s creations, a subsidiary company of Google. Their team is most famous for developing AlphaGo.
After becoming the world champion in the ancient board game, Go, it was usurped by its range of “Zero” AIs, those that can learn only by playing themselves. Unlike their predecessors, they require no human input whatsoever, and can master both Go and chess in a matter of days or even hours.
It looks like AutoML – the product of the Google Brain team – may turn out to be the company’s new star, however. An AI that can create its own AI sounds a little troubling, but it’s actually the opposite.
Say you need a complex computational or monitoring task to be run – say, climate modeling – but you can’t code. No problem: The AI could make a program for you with very little or no external input.