Earlier this month a self-driving shuttle in Las Vegas patiently waited as a delivery truck backed up, then backed up some more, then backed right into it. Inconveniently for the roboshuttle’s developer Navya, this happened within hours of the shuttle’s inauguration ceremony. The real problem is that the shuttle can’t learn from the incident the way a human would: immediately and without forgetting how to do everything else in the process.
The U.S. Defense Advanced Research Projects Agency (DARPA) is looking to change the way AI works through a program it calls L2M, or Lifelong Learning Machines. The agency is looking for systems that learn continuously, adapt to new tasks, and know what to learn and when. “We want the rigor of automation with the flexibility of the human,” says the program’s director Hava T. Siegelmann. The US $65-million program has already chosen 16 groups for 4-year projects, but according to Siegelmann there is still opportunity to propose 12- or 18-month projects.
AI’s big problem stems from the structure in use today. Neural networks are adaptable systems whose ability to learn comes from varying the strength of connections between its artificial neurons. Today these networks are trained on a set of data—images of cars and people for example. The strength of a network’s connections are then fixed, and the system goes out into the world to do its thing.
The problem comes when the AI encounters something it was never trained to recognize. Without retraining, the system would make the same mistake over and over again. But right now, AIs can’t really be retrained on the job. Trying to do so with today’s systems leads to a phenomenon called “catastrophic forgetting,” Siegelmann explained at the IEEE Rebooting Computing Conference. It’s a situation where learning the new item disrupts the knowledge of all the other things the system already knew how to do.
Even humans suffer some performance drop when they encounter something new, but we can recover while still performing a function. If you raise the net in a basketball game by 30-centimeters, players will miss most of the time at first, but as they continue playing they’ll learn to score at the new height. You don’t have to pull them off the court and teach them the entire game over again.
The 16 major grants went to two sets of groups. One set will have four years to develop systems that can continuously learn, adapt to new tasks and circumstances, and understand inputs according to what the system’s mission is (called “goal-driven perception). Another set will have four years to identify new mechanisms of lifelong learning—from biology or a physical science—and transfer that mechanism to an algorithm that improves AI.
Siegelmann says there are opportunities in both areas for shorter-term explorations. “It’s a good way to start working” with DARPA, she says. And it can form the basis for future programs. Contact Siegelmann here if you think you’ve got an idea that fits.