A new brain-inspired computer vision algorithm could help advance the state of AI—but it could also fill your in-box with spam.
Vicarious, a startup backed by both Elon Musk and Mark Zuckerberg, has published details of a new type of machine learning loosely inspired by neuroscience. The approach lets a computer learn to recognize visual information more efficiently and in a more human-like way—in that is doesn’t get confused when something is morphed slightly or partially occluded.
Vicarious, which has raised more than $150 million, has been talking about its technology for several years, and this is the first time it has published details (see “Inside Vicarious: The Secretive AI Startup Bringing Imagination to Computers”). In a paper published today in the journal Science, a team from Vicarious shows that its approach can beat Captcha—the squiggly, distorted text used to prevent bots from automatically registering e-mail accounts.
“We decided that it was better to be careful about releasing the details because of the obvious Web security implications,” says Dileep George, cofounder of Vicarious and the technical brains behind the company’s approach. George says most big companies have moved away from text-based Captchas, making it unlikely to be of much use to potential spammers.
Big strides have been made in computer vision over recent years using deep neural networks. These systems can learn to identify high-level concepts, like “dog” or “cat,” but they require huge amounts of data and are easily thrown off by things they haven’t seen before.
Vicarious developed what it calls a recursive cortical network (RCN) that can generalize beyond what it’s initially taught. Taking loose inspiration from neuroscience, an RCN is encoded with assumptions about visual information—like edges or curves—which it then uses to recognize inputs it hasn’t encountered in training. This means it can recognize a distorted “A” as long as it retains some of its key visual features.
The approach is similar to a technique used by researchers from MIT, CMU, and NYU to train a computer to recognize written characters from just one or two examples (see “This AI Algorithm Learns Simple Tasks as Fast as We Do”).
George says the technology developed by Vicarious could perhaps be used to help robots learn more efficiently. “We are working on several tasks in robotics,” he says. “Data efficiency and reasoning are very important when robots deal with unstructured environments.”