A team working on the Defense Advanced Research Projects Agency’s Lifelong Learning Machines program developed a new algorithm that allows machine learning systems the ability to continuously obtain new information and automatically adapt a task without requiring system resets.
DARPA announced Tuesday that researchers from the University of Southern California Viterbi School of Engineering in Los Angeles enabled an artificial intelligence-controlled robotic limb to teach itself how to walk and automatically recover after encountering an obstacle by using an artificial intelligence algorithm. The system was able to learn the task independently after five minutes of training.
Existing AI technologies require operators to overwrite their training set to learn new tasks. Such processes involve turning the technology offline and conducting another set of training.
“Current fixed methods underlying today’s smart systems will quickly give way to systems capable of learning in the field,” said Hava Siegelmann, a program manager at the DARPA Information Innovation Office. “We’re at a major moment of transition in the field of AI.”
She added the abilities to learn while in operation and apply learning to new circumstances would make AI systems safer than cars driven by people.
DARPA launched the L2M program for lifelong learning machines in 2017 to explore new ways to build next generation AI systems and replicate biological organisms to give the technologies new learning capability. The program supports 30 groups through grants and contracts.