Here’s Why This Cat-Spotting AI Is Different
In 2012, Google researchers found a whimsical way to show off the power of the company’s artificial intelligence technology: They trained computers to recognize cats in YouTube videos. The project took years to pull off and required 16,000 computer processors to analyze 10 million images. That type of AI, known as deep learning, now powers Amazon.com’s smart speaker Echo and Tesla’s self-driving cars. While such software can seem magical, it still typically requires thousands of computers to spend months scanning millions of data points.
Ben Vigoda, an MIT-trained computer scientist (and nephew of late actor Abe Vigoda), says he can cut out most of the grunt work and make AI projects doable for businesses without Google-level resources. His company, Gamalon Machine Intelligence, uses probability models to teach a computer to ID something like a cat in a few minutes by showing it just a few images. “You can run our software on a laptop, and it takes 100 times less horsepower to find an answer,” says Vigoda, who started Gamalon in 2013 but is showing its products to the public for the first time.
A typical deep-learning system needs to see tons of cats to learn all their variations—cats with every length of legs, tails, whiskers, and so on. Gamalon’s system gets a head start on this process with a new form of what’s known as probabilistic programming. Shown a cat with a long tail and one with a short tail, it can fill in the gaps. “The promise here is that you use data a lot more efficiently,” says Brenden Lake, a data science fellow at New York University.
Vigoda’s previous business, Lyric Semiconductor, hard-wired probability functions into computer chips. Lyric’s tech has found its way into Echo and cars, where it helps guess what people are saying, and into smartphones, where it improves the results of touches and swipes. With $4.5 million in funding from Silicon Valley investors and backing from the Defense Advanced Research Projects Agency, the Pentagon’s research and development arm, the Gamalon team has spent the past few years advancing probability technology to broaden its usefulness.
Academics have been working on probability-based AI systems for more than a decade, but few have made it to the real world. Microsoft’s TrueSkill, which ranks Xbox gamers, uses similar technology, as does pioneering startup Geometric Intelligence, acquired last year by Uber. Such systems have been tough to expand beyond very specific uses because humans still have to do the complex, time-consuming math to adjust their probability models. Gamalon’s big claim is that its software can adjust its own models, essentially reprogramming itself.
While Vigoda says his technology can be applied broadly, his company has started by focusing on a niche. Its first products, called Structure and Match, are designed to scour common databases and fix ambiguities, like different spellings for customer names and addresses. Communications-hardware maker Avaya, one of Gamalon’s handful of early customers, has been using the software—which costs $10,000 a month for analysis of each 100,000 rows of text in a database—for this kind of task.
Thousands of companies resell Avaya’s products and record customer data in different ways. To keep its databases effectively searchable, Avaya used to use people to pore through them for months at a time, turning “St.” into “Street” or “HP” into “Hewlett-Packard.” “With Gamalon, we were able to match 85 percent of the data in minutes instead of days,” says senior director Cary Gumbert. Unlike find-and-replace search tools, the software can recognize context, so it doesn’t turn, say, the “St.” in “St. Louis” into “Street.” “If we train it more and tell it where it has missed something,” Gumbert says, “it will only get better.”
The bottom line: Gamalon says it can cut AI training requirements from millions of photos and thousands of computers down to a few of each.