“When someone applies a machine learning algorithm, it’s hard to control its behavior,” said lead author Philip Thomas from the University of Massachusetts Amherst.
The key issue, according to Thomas and his fellow researchers, is not to try teaching computers to act morally, but to design the machine-learning algorithms so that they’re much easier for the (human) user to later add in constraints that make the AI safer and fairer.
The team came up with a framework they called ‘Seldonian’ algorithms, named after a character created by sci-fi writer Isaac Asimov, with which users can specify undesirable behavior as per their own needs.
Thomas gives the example of using their algorithm when controlling an insulin pump for diabetes treatment and specifying that undesirable behavior means dangerously low blood sugar. “Most algorithms don’t give you a way to put this type of constraint on behavior; it wasn’t included in early designs,” he explains.
In research published in Science, the team tested their Seldonian algorithm to predict grade point averages for 43,000 students in Brazil, and say it resulted in successfully overcoming several types of gender bias.
The researchers, who hail from UMass Amherst and Stanford in the US, and the Federal University of Rio Grande del Sol in Brazil, are calling on others in the field to continue their avenue of research for the development of more responsible machine learning.