Geoff Hinton, a 69-year-old AI researcher at Google stunned many in his community when he debuted Artificial Neural Networks back in 2012. The idea had been around for years, but Hinton made it possible and relatively simple to implement. In just five years, the tech has been built into almost all of our smartphones and devices, allowing easier speech transcription and facial recognition. Now, though, he’s reversed course and is looked to reinvent how AI is done.
Right now, neural networks are among the best techniques to manage computer vision. That’s the complex set of processes that allow computers to “see” and understand the world around them. While cameras can record, it takes a special kind of intelligence to interpret the dizzying complex array of patterns and shapes and everything else one gets with any image.
“I think the way we’re doing computer vision is just wrong,” Hinton told WIRED. “It works better than anything else at present, but that doesn’t mean it’s right.” That’s a bold claim, but coming from the man who practically invented modern AI, it carries some weight. And now, Hinton says he’s ready to push for the next thing — capsule networks.
Last week Hinton published a pair of papers that details the new approach. He says that these systems will make computers far, far better at understanding visual data. And with comparatively little development, Hinton’s new system does just as well as anything else we have, but they did so with half the mistakes of the next-best solution. The real kicker, though, is that this system requires far less raw data than something like Google’s DeepMind.
Neural networks are powerful, but they require thousands and sometimes millions of examples to study before it gets what a pencil is, for instance. In the era of big data, this isn’t too hard to achieve — we have trillions of photos, messages, videos, etc. all to feed into massive computer farms that can burn through that information in short order. It doesn’t however, lend itself well to new information, something humans can handle with relative ease.
Capsule networks work by creating clusters of digital neurons that each track a different element of the image. Stringing many capsules together gives you a system that much more closely resembles the way humans process and stores visual information. This allows it to understand, for instance, when you flip a pencil 180 degrees that it is still a pencil — without thousands of images from every conceivable angle to learn from.
Hinton has been working this idea for nearly 40 years, starting back in 1979. It also represents a major shift in how this kind of work is done. Recently, the trend has been to teach bots how to learn on their own from scratch. Contrary to what Locke may have you believe, humans aren’t blank slates when we’re born. So the approach doesn’t quite make sense if we’re talking about creating human-like intelligence. Instead, capsule networks are designed to mimic the basic instruction set many animals have from the start. You don’t need to teach a baby how to learn — it just does. That could be a huge difference in the coming years, but for now the tech is in its infancy.