The Fundamental Limits of Learning program aims to build a theoretical framework for teaching machines to generalize their existing knowledge to new situations.
“When we look at what’s happening with artificial intelligence, we see something that is very, very powerful, very valuable for military applications, but we also see a technology that is still quite fundamentally limited,” DARPA Director Arati Prabhakar said at the Atlantic Council on May 2.
Aiming to define those limits, a new Defense Advanced Research Projects Agency program will try to answer a singular question: “What are the fundamental limitations inherent in machine learning systems?”
Through a series of research areas of interest, the Fundamental Limits of Learning, or Fun LoL, program will assess the potential of focused investigations by developing, validating and applying a theoretical framework for learning, according to the recently released request for information.
The two primary research areas focus on articulating a general mathematical framework to measure learning and applying that framework to existing machine learning methods to characterize capabilities of current techniques.
Machines can master chess, Jeopardy! and Go, but “what’s lacking, however, is a fundamental theoretical framework for understanding the relationships among data, tasks, resources and measures of performance -- elements that would allow us to more efficiently teach tasks to machines and allow them to generalize their existing knowledge to new situations,” said DARPA Program Manager Reza Ghanadan said.
Because complex threats require machines to adapt and learn quickly, it is important that they be able to generalize creatively from previously learned concepts.
“If you slightly tweak a few rules of the game Go, for example, the machine won’t be able to generalize from what it already knows. Programmers would need to start from scratch and re-load a data set on the order of tens of millions of possible moves to account for the updated rules,” Ghandan said.
DARPA’s Fun LoL initiative is seeking information that could inform novel approaches to this problem. Technology areas that may be relevant include information theory, computer science theory, statistics, control theory, machine learning, AI and cognitive science.
Responses are due on June 7, 2016.