DARPA Is Funding Research to Find out Exactly What AI Is Thinking

When you were in high school, did you have to show your work in math class to get credit? The Defense Advanced Research Projects Agency (DARPA) wants artificial intelligence (AI) to do the same.

Recently, it announced the Artificial Intelligence Exploration (AIE) program, an initiative focused on fostering AI development that can understand and explain how it comes to conclusions.

Catalyzing AI Cognition

DARPA is a U.S. Department of Defense (DoD) agency concentrated on accelerating the creation of disruptive technologies. So it’s no surprise that they’re interested in making third-wave AI (the kind that can walk you through how it arrived at an answer) into a reality.

Most AI currently in use falls under the first-wave category. Basically, this type of AI can follow logical rules. Great examples of this are chess- or Go-playing AIs. Second-wave AI is a little more nuanced; they typically rely heavily on statistical learning to get results. An example of this would be image recognition.

A third-wave AI can do everything the first- and second-wave AIs can do, but it can also explain itself. For instance, let’s say you had AI that could identify an image of a dog. A third-wave version of this would be able to note that it came to this conclusion because of the animal’s tail, four legs, and other common characteristics of man’s best friend.

Funding the Future

Although it’s headquartered in Arlington, VA, far away from the San Francisco development community and the other AI hot spots, DARPA plans on tapping into all of this potential through funding. The way it works is the agency will release an “AIE Opportunity” notice that highlights a particular area of AI research that the military is interested in developing.

From there, researchers and developers can submit project proposals to DARPA. After review, the agency will award the winner with up to $1 million in funding. DARPA would like each initiative to be started within 90 days of the announcement and for winners to determine if the concept is even feasible within 18 months of commencement.

Thinking Differently

John Launchbury, Director of DARPA’s Information Innovation Office, believes third-wave AI systems will also require far smaller datasets than their first- and second- wave counterparts. For example, if an AI previously needed 100,000 images of handwriting to recognize it, a third-wave AI may only need one or two examples. Basically, it would break down how each letter is formed, and translate this context to identify other handwriting.

One of the biggest draws of third-wave AI is the fact that they’d ‘think’ rather than just come to calculated answers based on datasets and predetermined logical rules. Having AI that can reason and explain its abstract thoughts would have a huge impact on not only the field but how we all employ the technology in the future.