Framework for Avoiding "Adverse Outcomes" in Machine Learning

2019-12-11

Stanford University and University of Massachusetts Amherst researchers have introduced a framework for designing machine learning algorithms that make it easier for potential users to specify security and fairness constraints. According to the published paper, current machine learning algorithms "often exhibit undesirable behavior, ranging from various types of biases to causing economic losses or delaying medical diagnoses." What's worse is that the responsibility to avoid these pitfalls often falls on the algorithm's users rather than the designers, which is undoubtedly unfair. The researchers developed this new framework to allow users to easily constrain the behavior of algorithms without requiring extensive domain knowledge or additional data analysis. The authors state that this will shift the burden of ensuring the algorithm runs well from the users to the designers. As machine learning algorithms have a growing impact on society, the researchers believe it is crucial to establish safeguards against various "undesirable outcomes." Because if these outcomes can be mathematically defined, both users and algorithms can learn how to avoid them. The first author of the framework openly states that it is easier to ensure fairness, as it can avoid causing harm to numerous industries by generating "Seldonian algorithms." The framework is a tool that can guide researchers in creating algorithms that can be easily applied to real-world problems. "If I use a Seldonian algorithm to treat diabetes, then I can tell the machine, 'Do not make changes in the controller of the insulin pump that would increase the frequency of hypoglycemia,'" the researchers explain by way of example. "Most algorithms do not provide this type of application-specific behavioral constraint; that's where our inspiration comes from." The framework consists of three steps. First, defining the goals of the algorithm design process. Second, defining the interface that users will use. Finally, creating the algorithm. To demonstrate its feasibility, the researchers used the framework to design regression, classification, and reinforcement learning algorithms. The researchers tested and compared the framework on a student performance dataset, and the results showed that commonly used regression algorithms designed using standard ML methods exhibited discrimination against female students when fairness was not considered. In contrast, the researchers state that users can easily limit observed gender discrimination behavior using our Seldonistic regression algorithm. "Using algorithms designed with our framework is not just a replacement for existing ML algorithms in current applications," the researchers conclude in the paper. "We hope that it can pave the way for new applications that were previously considered too risky to use ML."

RECENT AI NEWS

RECENT AI TOOLS