Raytheon machine learning system will ‘know what it knows’
The Competency-Aware Machine Learning (CAML) program is focused on enabling learning systems to be aware of their own competency. Under this program, systems will be able to communicate the abilities they have learned, the conditions under which the abilities were learned, the strategies they recommend, and the situations for which those strategies can be used.
“The CAML system turns tools into partners,” says Ilana Heintz, principal investigator for CAML at Raytheon BBN Technologies. “It will understand the conditions where it makes decisions and communicate the reasons for those decisions.”
The system, says the company, will learn from a video game-like process: Instead of giving the system rules, the researchers will tell the system what choices it has in the game and what the ultimate goal is. By repeatedly playing the game, the system will learn the most effective ways to meet the goal. It will explain itself by recording the conditions and strategies it used to come up with successful outcomes.
“People need to understand an autonomous system’s skills and limitations to trust it with critical decisions,” says Heintz.
Once the system has developed these skills, the researchers will apply it to a simulated search and rescue mission. Users will create the conditions surrounding the mission, while the system will make recommendations and give users information about its competence in those particular conditions.
For example, says the company, the system might say, “In the rain, at night, I can distinguish between a person and an inanimate object with 90 percent accuracy, and I have done this over 1,000 times.”
According to DARPA, in addition to contracting with Raytheon, the CAML program – which began in February 2019 with an initial term of four years – will continue to seek expertise in machine learning, artificial intelligence, pattern recognition, knowledge representation and reasoning, autonomous system modeling, human-machine interface, and cognitive computing. The program is slated to involve three years of research and one year of technology demonstration.
DARPA looks to teach machines common sense
Self-learning robot a step closer to machine self-awareness
Self-taught Google AI achieves ‘superhuman’ chess proficiency in hours
Mission-level autonomy coming to DARPA small satellite network