A tiny group of artificial intelligence researchers say they’ve built solutions that are better than or nearly as good as the reigning global standards for determining how semantically related words are, in which categories images belong and how to analyze sentiment. For example, they are just .8 percent behind the winner, Google, in this year’s ImageNet competition, where teams compete on a computer vision challenge.
And all this only took them four months.
In their spare time, they raised $8 million from Khosla Ventures and Salesforce CEO Marc Benioff to launch a new company, called MetaMind.
What’s powering these breakthroughs? A branch of AI called deep learning that has had a big impact on the technology industry the last couple years because it is so good at assessing and labeling huge amounts of data. There are only a few longtime established deep learning researchers in the world, and the most prominent ones have been snatched up from their academic posts by Google, Facebook and Baidu.
Deep learning is not a novel technique, but it’s one that has recently been hugely successful for applications like image and speech recognition. And those things happen to be critically important products for big Internet companies.
One of the next generation of deep learning experts is Richard Socher, co-founder and CTO of MetaMind. He just got his PhD from Stanford this year. Now he’s out to hire and train a whole bunch of people to think like him, with the hope they can make big advances in applications for deep learning that aren’t locked up with those big Internet companies.
MetaMind will provide deep learning as a service to companies in industries like consumer goods, financial services and medicine. It wants to help radiologists identify cancer, insurers assess houses and nutritionists label food.
“We call it ‘drag, drop and learn,'” said MetaMind CEO Sven Strohband, who joined the company from Khosla Ventures. “All you need is a Web browser, and you can use deep learning technology.”
To that end, MetaMind has built some free Web apps that are open to the public. That includes a pre-built tool that will take in pictures of food and say what is on the plate. Or there’s another tool where, without writing any code, someone could develop an image recognition system that’s specifically trained to their friends’ faces.
To be sure, deep learning may have its moments, but it’s not the cure for every problem. Even Strohband would admit that. “We are trying to not be very abstract,” he said. “There are specific tasks, and we are trying to be the best in the world.”
MetaMind so far has 10 employees. What are the raw ingredients for a good deep learning researcher? Math, programming and smarts, said Socher.
This article originally appeared on Recode.net.