Neuronets poorly recognize black people and women

<pre>Neuronets poorly recognize black people and women

The face recognition system has already become quite dense in our lives. It finds application in scanners like Face ID and is even used to search criminals. But, according to the publication Science Daily, a group of researchers from the Massachusetts Institute of Technology (MIT) and Stanford, after studying several commercial AI algorithms for face recognition, revealed signs of sexism and racism in them.

The study revealed that when analyzing photographs white men, the algorithms incorrectly determined sex only in 0.8% of cases, and in the case of dark-skinned women, the error rate averaged about 30%. To reveal the regularities, the employee of MIT Joy Buolamvini collected 1200 photographs of different people (on which, however, mostly women with dark skin were imprinted). Then she turned to the Fitzpatrick scale, according to which there are 6 skin phototypes, where I is the lightest, and VI is the darkest.

As a result, it was found that for women with dark skin with a score of IV, V or VI on the Fitzpatrick scale, the percentage errors amounted to 20.8%, 34.5% and 34.7% respectively. Thus it was possible to establish that the darker the skin phototype – the greater the probability that the algorithm will make a mistake. When identifying men with fair skin, the percentage of errors did not exceed 0.8%. Moreover, Joy (who herself is African American) tried to analyze her own photos, and it turned out that the systems either do not recognize her face, or wrongly determine the gender.

In fact, such research can be treated differently. But throwing aside all the jokes and prejudices about tolerance, the research raises a very important question: are the face recognition algorithms really working properly? After all, with their help in the same China can already try to bring to criminal liability. In addition, the new survey is poorly correlated with the statement of one of the software companies that created software for neural networks and claimed that their software recognizes faces with accuracy in 97% of cases. And how do you think? Express your opinion in our new telegram chat.

Source link