Achieving Fairness in Machine Learning: New Approach by UNC Researchers

Machine learning algorithms have the potential to offer efficient and data-driven decisions, but their outputs have often been found to exacerbate existing biases and inequities rather than reduce them. Computer scientists have tried to implement fairness in these systems by designing various fairness principles, but a consistent application is yet to be made possible. Scholars from the University of North Carolina-Chapel Hill, however, offer an alternative approach; their research attaches explicit costs to a model’s classification errors and accounts for different types of costs to train the algorithm to consider context and asymmetries, rather than approach each decision uniformly. This alternative mindset is akin to economists who think in terms of cost-benefit analysis and utility functions. Recently, there has been growing societal concern over AI’s profound risks to society and humanity, prompting prominent industry leaders such as Geoffrey Hinton, Bill Gates, Elon Musk, and Steve Wozniak to call for urgent guardrails into place sooner rather than later.