DARPA and The Third Wave of AI

In February, U.S. Defense Advanced Research Projects Agency (DARPA) released the video “A DARPA Perspective on Artificial Intelligence,” featuring John Launchbury, the Director of DARPA’s Information Innovation Office. There is so much to love about this video.  

First, it was great to see how knowledge-based expert systems are still in use today despite constant talk in Silicon Valley about how expert systems “failed in the 80s.”  TurboTax is AI? Yep, that’s what it is; a bunch of hand coded logical steps to help you file   your taxes. These are the same logical steps you would go through if you were a tax expert.

Second, it was refreshing to see expectations tempered with respect to what Deep Learning can and cannot do. These systems have achieved truly incredible results over the past decade (AlphaGo was epic!), but seeing them for what they are (“spreadsheets on steroids”) reveals that we have a long way to go until all of our jobs are automated away. And, third, it was exciting to see DARPA set a higher bar for the AI community to work toward. This Third Wave, as they called it, is a fusion of first wave achievements (reasoning) and second wave achievements (learning and perceiving).

But why do we need a Third Wave? Well, despite the talk of “massive amounts of data” to train algorithms, in practice it is often very difficult to get your hands on enough clean data. New algorithms that can learn from small datasets are critical if we’re going to automate the majority of processes that we aspire to streamline. I find it interesting that people will anthropomorphize algorithms and in the same breath say that they need tens of thousands examples to train them to do a specific, one-off task that the average four-year-old could learn in five minutes. In school, I only needed to read a few textbooks to pass my exams and graduate; I didn’t need to read 10,000.  

For me, the most important reason for a Third Wave is explainability. Right now, Deep Learning systems can’t sufficiently explain to the average user how and why they arrived at a certain output. And, while expert systems can explain their outputs, the difficulty of manually coding all of the possible logical steps of every problem that could ever be encountered is simply untenable. But as I go through life, encounter new problems, and decide to take a certain action, I can almost always tell you why I did it. Our goal for the AI community in this Third Wave should be to enable the same capability in machines.

Ryan Welsh
CEO, Kyndi