Earlier this week, Stephen Wolfram unveiled a website that automatically identifies digital images. Drop in a photo of, say, a Tesla coil, and the site will tell you it's a Tesla coil.
Like so much that emerges from Wolfram Research---the eponymous software company operated by the British computer scientist, physicist, entrepreneur, and all-around free thinker---the site is a good time. It gets things right about as often as it gets them wrong. And, taken alongside Wolfram's typically expansive blog post detailing the project, it'll get you thinking about the future of artificial intelligence.
But in this case, Wolfram's demo also represents an enormous shift in AI that's happening right now. His tool is based on what are called "convolutional neural nets," vast networks of computer processors that attempt to mimic the networks of neurons in the human brain. As Wolfram points out, the neural net is a very old idea, dating back as many as six decades. But after years on the fringes of computer science---with many saying it would never work---this idea is now driving everything from Facebook photo recognition to Google voice recognition to Skype language translation.
"More and more companies are taking this kind of work very seriously," says David Luan, the founder of a neural networking outfit called Dextro.
Wolfram's new site shows that such AI is also readily available to software makers outside the big Internet giants---at least to a certain extent. The site is really just a demonstration of the latest edition to the Wolfram Language, the general-purpose programming language offered by Wolfram and company. Using the language, Wolfram says, any developer can build image recognition into their own application, tapping into a large cluster of machines operated by the company.
Other companies are doing similar work. An outfit called Metamind offers tools for driving your own applications with neural nets. Dextro offers neural-net-based tools that identify images in videos. And because many "deep learning" algorithms are available as open source software, even independent coders can run their own neural nets.
As Wolfram's demo shows, these techniques are still evolving. But it's now clear that neural nets work quite well, besting humans in some cases. They can reliably identity images and recognize speech and translate languages and more. Wolfram's demo shows that too.
This is particularly remarkable, Wolfram says, because the neutral net idea was presumed dead for so many years. "I don't know of any other technology where people were trying to do something that long ago and it finally worked," he says.
What changed is the amount of computing power we have at our disposal. We can now run these systems across dozens, hundreds, even thousands of high-powered processors. Much like Facebook and Google, Wolfram and company trained their image recognition model on a cluster of machines equipped with graphics processing units, or GPUs, low-cost chips suited to the kinds of calculations that drive neural nets. "The reason this has finally worked is not some great breakthrough," he says. "The reason is that we can now make systems that are big enough."
In some cases, even today's small systems are big enough. Says Yann LeCun, the head of Facebook's new artificial intelligence lab: "Any smart kid with a GPU-equipped PC can do this with open source tools in his parents' basement."
