Dark data: artificial intelligence learns to identify homosexuals in the photo

We have become to rely on machine learning systems in everything from creating playlists to control cars, but like any tool, they can be used in a dangerous and unethical purposes, even by accident. Today's illustration of this fact — a new article by researchers from Stanford who created a system of machine learning, which, they claim, can tell you several photo... gay in front of you. Or not.

Obviously, this study is surprising and disheartening in equal measure. Besides the fact that vulnerable group of people once again will be subject to a systematic invasion of privacy and intimate themes, this directly affects the egalitarian notion that we cannot (and should not) judge a person by his appearance, and does not have the right to guess the sexual orientation of one or two images. Absurd? The system accuracy reported in the work, however, leaves no room for doubt: it is now not only possible, but done.

The System picks up the emanations much more subtle than capable of vosprinimat people see the clue, which you do not even know. And it shows as intended that the privacy, private life in the era of ubiquitous computer vision can be simply destroyed.

Before proceeding to the discussion of the system itself, it is necessary to clarify that the study was conducted with good intentions. The questions, of course, there are countless. Most relevant, perhaps, were the remarks of Michal Kosinski and Elon Wong as to why the work actually saw the light:

"We were very concerned about these results and spent a lot of time to do them public. We didn't want to expose people to those risks about which they were warned. The ability to control when and to whom to disclose their sexual orientation, is crucial not only for human welfare but also for his safety."

"We thought that the legislators and the LGBTQ community needs to be aware of the risks faced. We did not create a tool of invasion of privacy, but rather showed that the simple and widely used methods pose a serious threat to life."

Definitely, this is just one of many systematic attempts to extract secret information such as sexuality, emotional state, or medical records. But in this case still more serious for several reasons.

the

to See that it is impossible (and not necessary)

In an article published in the Journal of Personality and Social Psychology, describes in detail quite a common approach based on reinforcement learning, the task of which is determination of the possibility to identify people as gay or not in their faces. Note: work is still in the form of a draft.

Using a database with pictures of faces (which were taken from Dating websites with open data), the scientists gathered 35 326 pictures 14 776 people, among which were equally represented homosexual men and women. Their facial features were extracted and calculated from the shape of the nose and eyebrows to facial hair and facial expressions.

The neural Network deep learning ate all of these features, noticing which are usually associated with a certain sexual orientation. Scientists are not "seeded" it with preconceptions about how usually are people with normal or homosexual orientation; the system simply assigned features and correlated them with the sexual preferences.

These patterns can be accessed to allow the computer to guess the sexual preferences of the person and it turned out that the AI system is much better at handling this task than humans. The algorithm was able to correctly determine the orientation of 91% of the men and 83% women, two at a time. The people who provided those images were right in 61% and 54% of cases, respectively — that is, a coin toss would give the same result.

Variations between the four groups described in the second article; aside from the obvious behavioral differences, like courting the same group or a similar makeup, the General trend was in the presence of "female" traits in gay men and "masculine" traits in lesbians.

the figure shows, where features indicating sexual orientation

It Should be noted that the accuracy of the system was achieved in ideal conditions the choice between two people, one of whom is homosexual. When the system was evaluated in a group of 1,000 individuals, of which only 7% belonged to homosexual people (for more representativeness is the actual percentage of the population), the results were not the best. In the best case a sample of 10 individuals were recognized with 90% accuracy.

There is also a real possibility of bias in the system due to the data: first, they included only young white Americans of the type of man/woman and homosexual/normal.

"Despite our attempts to obtain a more diverse sample, we were limited to the study of white participants from the United States. Because of prejudice against homosexuals and the use of online Dating sites are unevenly distributed among different ethnic groups, we were unable to find a sufficient number of non-white homosexual members."

Although scientists suggest that other ethnic groups are likely to have similar facial patterns and the system will also be effective to define, in fact it is necessary to confirm and not to assume.

You can Also argue that the classifier is collected, the characteristic features of people on Dating sites, not without the help of the mechanism of self-selection, for example, American men with a normal sexual orientation can deliberately avoided appearance, which would show that they can be homosexual.

Of Course, the real effectiveness of this system reasonably questioned, since it is trained based on the limited amount of data and mostly effective examples from this pool. Further research will be necessary. However, it seems short-sighted to assume that the running system is impossible.

the

the truth

Many argue that this technology is no better than "phrenology 2.0", shell when unrelated physical traits with complex internal processes, and that is understandable, but not true.

"Physiognomy is now widely and rightly rejected as a combination of superstition and racism disguised as science," the scientists write in the introduction to the work (physiognomy judging by the facial features, while the phrenologists studied the shape of the skull). But this pseudo science has failed because its supporters were moving in the opposite direction. They decided who was involved in "criminal cases" or "sexual deviations", and then determine what features are common to such people.

It Turned out that people just don't cope well with such tasks, we see people too much of myself and make instinctive conclusions. But this does not mean that the visible features and the physiology or psychology are not related. No one doubts that the machine can mark in people the traits that they do not see, from the tiny chemical traces of explosive shells and drugs to the subtle patterns on the MRI scans, which tell of a beginning cancer.

From the article it follows that the computer vision system chooses the same subtle patterns, both artificial and natural, but in the first place; the scientists tried to focus on traits that cannot be changed easily. This supports the hypothesis that sexuality is biologically determined, perhaps under the influence of various factors on the development of the fetus, which can result in small physiological differences and hormonal patterns of the appeal. The observed system patterns are similar to those predicted by the theory of prenatal hormones.

This is the starting point of a broader social and philosophical debates about love, sexuality, determinism and intelligence emotions and personalities to simple biological characteristics. But we won't. Better touch on other serious issues regarding confidentiality.

the

the Dangers and limitations

The effects of the automatic determination of the orientation of the man on several photos of the huge, first of all and most importantly, because put people of the LGBTQ groups at risk throughout the world in those places, where they remain oppressed minority. The potential abuse of such a system is huge and some may rightfully disagree with the decision of the researchers about its creation and documentation.

But similarly you could argue that it is better to be able to open it, to implement countermeasures or otherwise to allow people to prepare for it. The difference is that you can hide the information you post online; you can't control who sees your face. It is regulated by the society in which we live, and concealed from common view is less and less.

"Even the best laws and technology to protect privacy, in our opinion, doomed to failure," the scientists write. "The digital environment is very difficult for the police; the data can be easily transferred across the border, to steal or to record without consent".

This issue is infinitely complex, but it demonstrates the amazing capabilities of artificial intelligence. The consequences of such technologies, as usual, ambiguous, but very interesting for the future of humanity and the individual. Scientists themselves are not particularly optimistic about the use of such technologies in the future, but are realistic in an era when technology, which it is expected the plan of salvation, most likely will play the villain.

"We believe that further erosion of private life is inevitable, and the safety of gays and other menshinstv does not depend on the right to privacy and respect for human rights and tolerance of society and the state", they write. "...