MIT hopes to automatically 'de-bias' face detection AI
Its algorithm promises wider representation with minimal effort.
There have been efforts to fight racist biases in face detection systems through better training data, but that usually involves a human manually supplying the new material. MIT's CSAIL might have a better approach. It's developing an algorithm that automatically 'de-biases' the training material for face detection AI, ensuring that it accommodates a wider range of humans. The code can scan a data set, understand the set's biases, and promptly resample it to ensure better representation for people regardless of skin color.
The technology won't necessarily iron out all biases, but the results can be significant. In testing, MIT's system reduced "categorical bias" by 60 percent without affecting the precision. It also promises to save time, especially for larger data collections that are time-consuming.
You might not see this approach used in practice for a while. However, it could become crucial as police and companies rely more and more on face detection. Biased face recognition can not only make it harder to use certain devices, but produce false positives that lead to arrests of innocent people. Automatic bias removal could alleviate some of that worry -- you might not like face-detecting AI in the first place, but this would at least spare you from discrimination when the technology comes into play.
