With $8M and star team, MetaMind does deep learning for enterprise

A Palo Alto startup called MetaMind launched on Friday promising to help enterprises use deep learning to analyze their images, text and other data. The company has raised $8 million from Khosla Ventures and Marc Benioff, and and Khosla operating partner and CTO Sven Strohband is its co-founder and CEO. He’s joined by co-founder and CTO Richard Socher — a frequently published researcher — and a small team of other data scientists.

Natural language processing expert Chris Manning of Stanford and Yoshua Bengio of the University of Montreal, considered one of the handful of deep learning masters, are MetaMind’s advisers.

Rather than trying to help companies deploy and train their own deep neural networks and artificial intelligence systems, as some other startups are doing, MetaMind is providing simple interfaces for predetermined tasks. Strohband thinks a lot of users will ultimately care less about the technology underneath and more about what it can do for them.

“I think people, in the end, are trying to solve a problem,” he said.

Sven Strohband (second from left) at Structure Data 2014.
Sven Strohband (second from left) at Structure Data 2014.

Right now, there are several tools (what the company calls “smart modules”) for computer vision — including image, localization and segmentation — as well as for language. The latter, where much of Socher’s research has focused, includes modules for text classification, sentiment analysis and question-answering, among other things. (MetaMind incorporates a faster, more accurate version of the etcML text-analysis service that Socher helped create while pursuing a Ph.D. at Stanford.)

During a briefing on MetaMind, Socher demonstrated a capability that merges language and vision and that’s similar, inversely, to a spate of recent work from Google, Stanford and elsewhere around automatically generating detailed captions for images. When he typed in phrases such as “birds on water” or “horse with bald man,” the application surfaced pictures fitting those descriptions and even clustered them based on how similar they are.

Testing out MetaMind's sentiment analysis for Twitter.
Testing out MetaMind’s sentiment analysis for Twitter

Socher and Strohband claim MetaMind’s accuracy in language and vision tasks is comparable to, if not better than, previous systems that have won competitions in those fields. Where applicable, the company’s website shows these comparisons.

MetaMind is also working on modules for reasoning over databases, claiming the ability to automatically fill in missing values and predict column headings. Demo versions of several of these features are available on the company’s website, including a couple that let users import their own text or images and train their own classifiers. Socher calls this “drag-and-drop deep learning.”

The bare image-training interface.
The bare image-training interface

On the surface, the MetaMind service seems similar to those of a couple other deep-learning-based startups, including computer-vision specialist Clarifai but especially AlchemyAPI, which is rapidly expanding its collection of services. If there’s a big difference on the product side right now, it’s that AlchemyAPI has been around for years and has a fairly standard API-based cloud service, and a business model that seems to work for it.

After being trained on 5 pics of chocolate chip cookies and five pics of oatmeal raisin cookies, I tested it on this one.
After being trained on five pictures of chocolate chip cookies and five pictures of oatmeal raisin cookies, I tested it on this one.

MetaMind is only four months old, but Strohband said the company plans to keep expanding its capabilities and become a general-purpose artificial intelligence platform. It intends to make money by licensing its modules to enterprise users along with commercial support. However, it does offer some free tools and an API in order to get the technology in front of a lot of users to gin up excitement and learn from what they’re doing.

“Making these tools so easy to use will open up a lot of interesting use cases,” Socher said.

Asked about the prospect of acquiring skilled researchers and engineers in a field where hiring is notoriously difficult — and in a geography, Palo Alto, where companies like [company]Google[/company] and [company]Facebook[/company] are stockpiling AI experts — Socher suggested it’s not quite as hard as it might seem. Companies like MetaMind just need to look a little outside the box.

“If [someone is] incredibly good at applied math programming … I can teach that person a lot about deep learning in a very short amount of time,” he said.

He thinks another important element, if MetaMind is to be successful, will be for him to continue doing his own research so the company can develop its own techniques and remain on the cutting edge. That’s increasingly difficult in the world of deep learning and neural network research, where large companies are spending hundreds of millions of dollars, universities are doubling down and new papers are published seemingly daily.

“If you rest a little on your laurels here,” Strohband said, “this field moves so fast [you’ll get left behind].”