Researchers at the Massachusetts Institute of Technology's computer science and artificial intelligence laboratory in Cambridge are working to create an AI algorithm that can "de-bias" other AI systems.
"We've learned in recent years that AI systems can be unfair, which is dangerous when they're increasingly being used to do everything from predict crime to determine what news we consume," reads an article from MIT CSAIL. "Last year's study showing the racism of face-recognition algorithms demonstrated a fundamental truth about AI: If you train [AI] with biased data, you'll get biased results."
Three notes on how the research team is trying to fix this issue:
1. The new AI algorithm is designed to "de-bias" data by resampling it to be more balanced. The algorithm learns both the underlying structure of the training data and the task that the original system was trained to do — such as face detection for a facial-recognition system — which allows it to identify and subsequently minimize existing biases.
2. Unlike existing approaches to resampling data that tend to require some level of human input to define specific biases, the MIT team's algorithm reviews a data set on its own to learn what biases are "hidden inside it," and automatically resamples it to be more fair, according to the article. This makes the algorithm particularly useful for larger data sets that are difficult to vet manually.
3. In tests, the algorithm decreased "categorical bias" by more than 60 percent compared to existing state-of-the-art facial-detection models, according to the article.
"Rectifying these issues is especially important as we start to see these kinds of algorithms being used in security, law enforcement and other domains," PhD student Alexander Amini, who was co-lead author on a related paper, said in the article. The paper was presented in January at the Conference on Artificial Intelligence, Ethics and Society in Honolulu.