Microsoft is updating its facial recognition tech to work better for everyone. The company identified diversity issues with facial recognition technologies throughout the industry, stating that the systems are usually less accurate on darker skin tones, and performed worse for females with darker skin. So what’s the reason behind all these anomalies?
Microsoft is blaming it on the training data used to train its machine learning models for the facial recognition tech. The company says the training data used to train these models aren’t diverse enough, meaning they mostly include males with light skin tones, less of people with darker skin tones and less of females with dark skins. “Collecting more data that captures the diversity of our world and being careful about how to measure performance are important steps toward mitigating these issues,” Microsoft Research’s Ece Kamar said in a statement.
To mitigate the issue, Microsoft expanded its dataset to include more diversity, with an expanded dataset that includes things like different hairstyles, jewellery, eyewear, as well as the skin tones. With the new dataset, Microsoft has been able to reduce the error rates for all genders with a dark skin by up to 20 times, and the error rates were reduced by 9 times for all women.