American researchers testing the accuracy of face recognition technology introduced by IBM and Microsoft found that the services tested were more accurate in white color and less accurate in dark color people.The report states that the reason for this is that The original training algorithm has less black population. Experts also call for industry to be more transparent in the current identification of technical accuracy information. According to Wired, the researchers tested the facial analysis services on Microsoft and IBM to identify gender on photos. It is found that the accuracy of the company's algorithm in identifying men with white complexion is close to 100%. However, when analyzing women with dark complexion, they often make mistakes. This is due to the fact that machine learning techniques are used to create darker skin tone in the training data of facial analysis algorithms. Less. In the past, Google’s photo classification service used to label blacks as chimpanzees and it made the outside world quite different. Therefore, how to ensure that the machine learning system deployed in consumer products, commercial systems, and government programs is flawless is already artificial intelligence (AI). Important topics for discussion in the field. A report from Georgetown University in 2016 points out that it has been part of the FBI and local police. Of the face recognition technology is low on African Americans.In this new study, MIT researcher Joy Buolamwini and Microsoft researcher Timnit Gebru invested in the recognition system 1,270 Europe Photographs with African parliamentarians, but also with the Fitzpatrick scale classification system. This photo gallery was used to test Microsoft's Face+ cloud recognition service for face recognition by the Megvii company, a startup company in mainland China. The three services in the gender detection function and found that these three services in the face of men than women, the accuracy of higher than the white color is better than the dark skin, all services for women with dark skin, the accuracy is Microsoft did not respond correctly to white-skinned men, with an IBM error rate of 0.3%, but Microsoft's error rate was 21% for Blacks and 35% for both IBM and Face ++. On-demand machine learning algorithm services have become a hot area for large companies to compete. Microsoft, IBM, Google, and Amazon all advertise that the cloud is suitable for tasks such as grammatical analysis of imagery and textual meanings. Such as sports, healthcare and manufacturing, but at the same time, customers may have to be forced to accept the limitations Pivothead, a startup that develops smart glasses for the visually impaired, is one of the customers who leverage Microsoft's AI services, However, Microsoft's technical papers point out that gender recognition and other facial features such as emotion and age are currently experimental and not entirely accurate, and DJ Patil, former chief information scientist for President Obama, He pointed out that this report highlights that tech companies must ensure the importance of their machine learning systems for all types of humans, and that the industry must also be more open about the limitations of their services. Microsoft is currently working hard on the importance of machine learning ethics. As a leader, many of the company’s researchers are investing in the field and setting up an internal AI and Aethical Research and Research Committee. The unit participated in the audit in 2017 and found that Microsoft’s cloud service for analyzing facial expressions was used by children. Poor performance.
