Artificial-Intelligence Startup MetaMind’s Bright Idea: Give Its Software Away

MetaMind customizes its deep-learning software for businesses that want to learn faster

It may not seem like it, but artificial intelligence is getting closer to making the leap from the laboratory to the marketplace. For years, the big companies at work on A.I. have mostly kept their research to themselves, occasionally going public with a stunt to hint at their progress. IBM famously trained its Watson system to beat Jeopardy! champion Ken Jennings, Microsoft showed off a version of Skype that could translate a multilingual conversation in real time, and Google taught a network of 16,000 computers to identify a cat. All were important, if incremental, efforts to develop computer networks that can simulate the brain’s capacity for learning—though not on the level of movie menaces such as HAL or Skynet.

Richard Socher is trying to bring A.I. tools to a larger audience. A Stanford Ph.D., Socher in 2009 helped create ImageNet, a set of benchmarks that A.I. researchers use to compare their image-recognition software in an annual online competition. While finishing his doctorate this summer, Socher and Sven Strohband, the chief technology officer at venture fund Khosla Ventures, co-founded MetaMind, a startup that has created a hybrid of text- and photo-recognition software. The software can teach itself to remember words and images so that it can analyze new data, from medical scans to 10-Q filings. A basic version is available free online, but the company charges clients to customize it, says Strohband, MetaMind’s chief executive officer.

Give MetaMind’s software a bit of text or a few sample images, and it can compare two sentences for similarities, categorize tweets on a given subject as positive or negative, or distinguish chocolate chip cookies from oatmeal-raisin based on a handful of photos. The 10-person startup also writes new code for its clients. MetaMind taught its software to study mammograms and incorporate doctors’ reports, to identify evidence of breast cancer, for a radiology company. For a consumer electronics company, it wrote an algorithm that could recognize foods based on cell phone photos; Strohband says it’s working on ways to count calories looking at images.

MetaMind is also working to help a financial-services firm evaluate the risk of a stock selloff by scanning corporate financial disclosures, and to help a customer-retention company identify online interactions that suggest when a retailer needs to reach out to an angry customer, says Strohband, who wouldn’t disclose names of clients or discuss fees, except to say that price varies by project. The company has put free A.I. online for two reasons, Socher says: to attract clients for its customization business and to feed data to its servers so the software can keep learning. “As we explore and observe people using the platform, I think the platform will get smarter and smarter,” he says.

The software’s advantage, Socher says, is that it can be more easily customized for clients than competing programs. Most A.I. programs can’t scan written and visual media at once, he says. (IBM’s Watson was built for text; rival startups have also focused on specific industry uses.) Matthew Zeiler, the CEO of image-focused A.I. startup Clarifai in New York, says his company has been most focused on “real-world applications” such as tracking brands in social media photos or scanning security footage for incidents. Like Strohband, he wouldn’t name clients.

Socher moved to the U.S. from Germany in 2008 to study deep learning at Stanford after earning a bachelor’s degree from Leipzig University and a master’s from Saarland University. (His thesis in language processing carried the title “A Learning-Based Hierarchical Model for Vessel Segmentation.”) The son of engineering and chemistry Ph.D.s, he worked at Stanford with renowned computer scientist Andrew Ng, won best paper at the International Conference on Machine Learning in 2011, and received a Microsoft Research fellowship the following year. By the time he defended his Ph.D. thesis, he’d lectured A.I. researchers at Google, Microsoft, IBM, and Facebook. “There is no doubt that he is one of the young stars of the field from an academic point of view,” says Yoshua Bengio, a professor at the University of Montreal who’s advising MetaMind.

Instead of accepting offers for academic positions or jobs with other companies, Socher says, he co-founded MetaMind because he saw a huge demand for A.I.-driven data analysis that wasn’t being addressed in the market. On Dec. 5 the company announced that it had received $8 million in funding from Khosla (surprise) and Marc Benioff, the CEO of Salesforce, which makes sales and marketing analysis software. Strohband says he expects MetaMind to turn a profit next year.

MetaMind could have trouble finding a steady supply of customers who are savvy enough to use its software but unschooled enough to need its consulting services, says Andrew Maas, a co-founder of Roam Analytics, which uses machine-learning tools to build sales software but doesn’t compete with Socher’s startup. “The end user has to be smart enough that they’re going to be able to give you the data and set up the problem in a format that is solvable by the techniques we have,” Maas says. “But at the same time, the user has to be not smart enough or not experienced enough to use the open-source tools and run it themselves.”

By at least one standard—the ImageNet benchmarks Socher designed—MetaMind is near the front of the pack. Its software correctly classifies objects 92.4 percent of the time; based on the results from this year’s ImageNet competition, that’s second only to Google.

    Before it's here, it's on the Bloomberg Terminal.