The Next Big Thing You Missed: The Quest to Give Computers the Power of Imagination

Vicarious wants to make artificial intelligence infinitely more intelligent. Here's why tech giants like Mark Zuckerberg and Jeff Bezos have given the startup millions of dollars to do it.
Photo Courtesy of Vicarious
The team behind Vicarious, a new kind of artificial intelligence company.Photo: Courtesy of Vicarious

Imagine you want to teach a child the meaning of the word "table." You'd probably point to a few examples in your home--a round wooden kitchen table, a square plastic kiddie table, a massive rectangular dining room table. "This is a table," you'd say, three or four times, and after a while, the child would start identifying other tables, regardless of their shape, size, or color.

If you want to teach a modern computer the meaning of the word table, the process isn't much different--except that you can't stop at three or four examples. When dealing with a machine, you have to show it millions of tables before it can accurately identify a table on its own. This typically is called artificial intelligence, but D. Scott Phoenix doesn't see it that way. "It's actually pretty dumb," says Phoenix, the founder of the three-year-old Silicon Valley startup Vicarious. Intelligence, he explains, is "being able to deduce something from very few examples."

With Vicarious, Phoenix wants to create a form of artificial intelligence worthy of the name. With co-founder Dileep George, a respected AI and neuroscience researcher, Phoenix and a team of about 10 others are working to build a system that learns faster than other AI methods by taking advantage of what may be the brain's most important asset: imagination. Imagination is what lets a child figure out that a wooden table is still a table, even if he's only ever seen a plastic table, and this imaginative leap involves a series of complex neural connections that link shape, texture, color, and depth. A human brain can make those connections naturally, but even today's best artificially intelligent computers and online services cannot. That's the problem that Vicarious is trying to fix.

>Imagination is what lets a child figure out that a wooden table is still a table, even if he's only ever seen a plastic table, and this imaginative leap involves a series of complex neural connections that link shape, texture, color, and depth.

If Vicarious can pull that off, it could change the way we interact with computers, and how computers interact with the world, for decades. That potential hasn't gone unnoticed by the biggest names in tech. Just last month, Vicarious closed a $40 million funding round, which included investments from the likes of Amazon founder Jeff Bezos, Facebook founder Mark Zuckerberg, big-name venture capitalist Peter Thiel, and, reportedly, Tesla founder Elon Musk (Phoenix declines to confirm Musk's investment). That brings its funding to around $60 million.

The startup is part of a much larger effort to create brain-like systems that are more capable of not only identifying objects but understanding natural language. Drawing on a burgeoning academic field known as "deep learning," Google and Facebook are at the forefront of this movement. Google, for instance, is already using these techniques to process voice commands on Android phones. But Vicarious aims to eclipse even deep learning, which involves processing massive amounts of data on the fly.

Phoenix says he went the startup route for a reason. Typically, advancements in artificial intelligence have occurred within an academic setting, where researchers are working on a nine- to 18-month cycle, from funding grant to final publication. Phoenix believes this is unsuited to a moonshot like Vicarious. "If you're going to do something drastically different, it's a big risk to your career," he says of academic research. "If you spend two years working on a new approach, and it doesn't work better than existing approaches, you've effectively wasted those two years and can't publish a paper." So, in the late aughts, he began researching artificial intelligence methods on his own, and that's how he met George. The two founded Vicarious in 2010.

The best way to understand the company's technology is to compare it with other AI methods, including deep learning. Phoenix admits he's exaggerating when he calls deep learning "dumb." Driven by people like The University of Toronto's Geoffrey Hinton and NYU's Yann LeCun, deep learning has been percolating for decades and has captured the attention of Microsoft and Netflix as well as Google and Facebook. But for Phoenix, it's a limited technology. The simplest way to understand how deep learning works, he says, is to imagine a database that contains millions of images. A deep learning system must sift through all these images before it's about to identify something it's looking at. It's a one-way connection.

The human brain doesn't work that way. It makes connections in both directions, constantly feeding new information back to build upon what it's already learned. And that's what Vicarious wants to mimic. One of the more noteworthy breakthroughs of Vicarious' system, for instance, has been its ability to solve CAPTCHA codes, those distorted letter combinations that help prove you're human online. Because deep learning systems have never seen letters in exactly that shape or oriented in exactly that way before, they can't tell where one letter ends and the next begins. But because Vicarious' technology can imagine new positions and orientations, it can parse out individual letters to crack the code.

"This, by itself, is impressive," says Abdel-rahman Mohamed, who received his Ph.D. in computer science at the University of Toronto and studied alongside Hinton, the man at the center of the deep learning movement. What Mohamed has yet to see proof of, however, is how Vicarious can extend this work to more ambitious applications. "It works for CAPTCHA, but can it work for reading handwriting?" he asks.

Though Phoenix remains secretive about how Vicarious' technology actually works, above is a visualization of the inner workings of its software.

Image: Courtesy of Vicarious

Phoenix and Vicarious are rather secretive about their technology. But he says decoding CAPTCHAs is just a beginning. The company has no actual product in the works and no profit in mind, but aims to create a system that does much more than enhance a social network or search engine. True success will come when the system can identify a tumor on an X-ray, spot a default in a manufactured part, or enable a robot to autonomously navigate a kitchen.

Is any of this doable? As it stands, we have no way of telling. "I will not say they can't do it," says Mohamed, "and I won't be a believer until I see some strong evidence, either." But Mohamed places a great deal of faith in the fact people like Bezos and Zuckerberg have invested in Vicarious. And they aren't the only ones. Brian Singerman, general partner at Founders Fund, says the primary reason his fund invested in Vicarious was because of the team. "The difference between tech investing and product investing is you can't predict when a technological breakthrough will occur," he says. "We're comfortable with that, as long as we believe the team and where they're trying to go."

According to Phoenix, Bezos was tougher to pitch, but ended up investing because of Vicarious's longterm potential to make robots--be they drones or factory bots--vastly more intelligent. And while Phoenix insists he hasn't spoken to Zuckerberg about Vicarious' potential applications at Facebook, which has its own AI lab, it's not hard to imagine how Vicarious' technology could make everything from ad recommendations to photo identification more efficient. "AI is the one problem that, if you solve it, you've solved many problems, maybe all problems," Phoenix says. But he adds a caveat. You must solve it in the right way.