Google Says Its AI Can Program AI Better Than Humans Can

Google’s AutoML programme, announced earlier this year, has now produced machine-learning models that outperform hand-built ones

Google’s AutoML programme, announced earlier this year, has now produced artificial intelligence software capable of coding more AI tools that are better in some ways than what a human could come up with, in a development that could help the technology spread more broadly, the company said.

Google chief executive Sundari Pichai
Google chief executive Sundari Pichai

One of AI’s bottlenecks is that there are so few specialists capable of working on its development, and Google took aim at this problem with AutoML, announced in May.

Neural nets

The programme focuses on a particular kind of machine learning model known as a neural net, based loosely on the functioning of the human brain.

Google chief executive Sundar Pichai said at the time of the announcement that the programme aimed to design neural nets capable of designing neural nets.

He said at the time that the manual design of neural nets is “extremely time intensive” and requires expertise even more specialised than AI in general.

“We hope AutoML will take an ability that a few PhDs have today and will make it possible in three to five years for hundreds of thousands of developers to design new neural nets for their particular needs,” Pichai said in a May blog post.

Now the programme has shown its first results, with a system designed by a neural net scoring a record 82 percent at categorising images by their content, Google said.

On distinguishing the locations of multiple objects in an image, a more difficult task, the tool scored 43 percent, compared to 39 percent for the best equivalent built by humans.

Talent shortage

Google is one of the biggest corporate investors in AI, and under its “AI first” strategy uses the technology internally to run more efficiently and build new products.

Industry observers believe AI is close to a tipping point that could soon see it automate entire job categories, including those associated with highly trained humans in areas such as healthcare and finance.

But as the government’s review of the UK’s AI industry highlighted this week, the technology’s development depends upon specialised researchers who are in short supply, even at companies such as Google.

Programmes such as AutoML could help solve that problem by making researchers’ efforts go farther, Google said.

Google competes in the area with companies including IBM and Microsoft.

It has bolstered its own efforts by acquiring third-party leaders in the industry, including UK-based AI firm DeepMind, which it bought in 2014.

Put your knowledge of artificial intelligence (AI) to the test. Try our quiz!