WHY THIS MATTERS IN BRIEF
It is inevitable that one day the majority, if not all, new AI’s will be created by AI’s and not humans, and that day might come sooner than we think.
Machine learning experts and data scientists are in increasingly short supply as organisations around the world, and in every industry, race to take advantage of the recent advances in Artificial Intelligence (AI). Now Google’s CEO, Sundar Pichai, says he believes that one solution to the skills shortage is to automate the creation of new AI’s by having machine learning software take over some of the heavy lifting.
In other words he believes that the best way around the problem is to get AI’s to design and create new AI’s, and while this might sound like the musings of an executive thinking through the problem and then coming up with a potential solution though the fact of the matter is that, on the one hand, Facebook’s AI’s have been creating new AI’s for almost a year now, Microsoft’s AI’s are designing and writing their own programs, with a great deal of success, and late last year, as I reported, Google’s engineers quietly announced they’d been working on the project for some time now.
One Google project in particular, for example, called AutoML, which was unveiled at Google’s annual I/O developer conference the other week, has already demonstrated its expertise at architecting and designing new AI’s that rival, and in many cases, beat the best work of human machine learning experts.
“This is a very exciting development,” said Pichai, “it could accelerate the whole [AI] field and help us tackle some of the most challenging problems we face today.”
Pichai particularly hopes the AutoML project, which targets the development of new deep learning platforms that are used in everything from speech and image recognition, language translation and robotics, can help lower the barriers to entry, democratise and open the field up to those developers who don’t have the same levels of expertise as some of their peers.
Deep learning teaches software to be smart by passing data through layers of maths that are loosely inspired by the brains own neural network architecture, and as a consequence choosing the right architecture to use for a neural network’s web of maths is a crucial part of making an AI that works – but it’s not easy to figure out.
“At the moment we do it by intuition,” says Quoc Le, a machine learning researcher at Google working on the AutoML project.
Last month, Le and his fellow researcher Barret Zoph, presented the results from their experiments where they’d tasked their new machine learning system with figuring out the best deep learning architecture to use to solve language and image recognition tasks. On the image task, their system rivalled the best architectures designed by human experts, and on the language task, well, it beat them…
Perhaps more significantly though it came up with radical new architectures that the researchers hadn’t thought of using before.
“In a sense it found something we didn’t know about,” says Le, “it’s striking.”
The notion of AI’s that learn and improve over time has been around for decades, but unlike yesteryear the power of today’s technology and AI’s are helping researchers craft breakthrough after breakthrough.
When asked the inevitable question, are Le and Zoph on track to put themselves out of a job, the pair laugh. Right now the new technique is too expensive to be widely used – the pair’s experiments tied up 800 powerful graphics processors for multiple weeks and racked up the kind of power bill that few companies could afford for speculative research.
Still, Google now has a larger team working on AutoML, and that includes teams who are working on trying to make it less resource intensive, and one day, sooner rather than later, we’ll inevitably see a new dawn when machine learning experts and data scientists become increasingly automated.