Over 30 years ago when I first started to program I was quickly fascinated by the potential for computers to become smart. I remember writing my own version of Eliza and buying the Winston’s book on Artificial Intelligence. It seemed like anything was possible and thinking machines would be just around the corner. Then followed nearly a quarter century of, well, disappointments.
Tasks that were simple for humans, such as figuring out what’s in a picture, seemed to defy all efforts by machines. When spam and fake accounts became a problem on the Internet we wound up being able to use simple images of slightly distorted numbers or letters to tell machines and humans apart in what became known as a CAPTCHA, short for Completely Automated Turing Challenge.
Alan Turing, the pioneer of theoretical computer science and cryptology, had proposed a test in which a human would be allowed to ask questions of multiple respondents. If the human couldn’t tell which respondent was a machine then that machine should be considered (at least somewhat) intelligent. You can think of a CAPTCHA as a miniature version of such a test, which is why sometimes they are accompanied by questions such as “Are you a human?”
But then suddenly a few years ago most CAPTCHAs stopped being effective. Machines were easily able to recognize the images as quickly as a human. And that turned out to be part of a much broader trend of breakthroughs in making machines smarter. For example, we went from first versions of self driving cars that barely made it one mile on a closed course in 2004 to self driving cars successfully navigating on crowded highways by 2012.
How did all of this happen? There are of course many contributing factors, such as much faster CPUs and better algorithms. But the most important breakthrough has been the ready availability of vast amounts of data from which machines can learn. As it turns out, not unlike humans, machines learn by seeing lots of examples of what something is and isn’t. This kind of “training data” is now everywhere and growing rapidly. For instance, every day millions of new images are categorized by humans who tag them with information about location and contents.
Still, until recently it was difficult for developers to access these breakthroughs. Much of it was happening in either academic research or at the largest companies. We are therefore excited to be investors in Clarifai, a New York based company that is making advanced image and video classification solutions available as an easy to use API.
With more and more visual content everywhere search and discoverability are becoming paramount. Marketplaces, in which products can be listed with images, are just one of many examples that can use Clarifai to offer a better experience. If you want to play around with the technology without programming you can just take an existing image or video and try out Clarifai’s classification demo. We look forward to seeing what people will build on top of Clarifai — if you have a small project in mind you may want to participate in the upcoming video hack day. You can also read more about the financing on the Clarifai blog.
PS Clarifai is hiring for many different roles!