New Algorithm trains AI to avoid bad behaviors

Robots, self-driving cars and other intelligent machines could become better-behaved thanks to a new way to help machine learning designers build AI applications with safeguards against specific, undesirable outcomes such as racial and gender bias. Artificial intelligence has moved into the commercial mainstream thanks to the growing prowess of machine learning algorithms that enable computers to train themselves to do things like drive cars, control robots or automate decision-making. Go to the web site to As robots, self-driving cars and other intelligent machines weave AI into everyday life, a new way of designing algorithms can help machine-learning developers build in safeguards against specific, undesirable outcomes like racial and gender bias, to help earn societal trust. But as AI starts handling sensitive tasks, such as helping pick which prisoners get bail, policy makers are insisting that computer scientists offer assurances that automated systems have been designed to minimize, if not completely avoid, unwanted outcomes such as excessive risk or racial and gender bias. A team led by researchers at Stanford and the University of Massachusetts Amherst published a paper Nov. 22 in Science suggesting how to provide such assurances. The paper outlines a new technique that translates a fuzzy goal, such as avoiding gender bias, into the precise mathematical criteria that would allow a machine-learning algorithm to train an AI application to avoid that behavior.