This article is from the public number:All Media (ID:quanmeipai), the original title is: “Face recognition behind the gender bias of the skin color, is the algorithm blind spot or the heart of the speculation? “, the title map comes from: Visual China.
A recent study by the Pew Research Center found that FacebookIn the news photos, men appear twice as often as women, and most of the pictures are about men.
Considering that 43% of adult American citizens currently access news via Facebook, the Pew Research Center uses machine vision to test the news photos posted on Facebook by 17 national news media between April and June 2018. Gender ratio situation. The test algorithm finally identified 53,067 people, of which 33% were women and 67% were men. But in real life, the sex ratio of the US population is roughly balanced.
So, who is “distorting” the two sexes?
Why sometimes, in the eyes of the algorithm, are you in a fuzzy area where you can be a man?
Further, what are the biases beyond gender?
What can we do to deal with this situation?
Gender imbalance under face recognition
Pui’s report points out that in different types of news reports on Facebook, women’s “presence” in pictures is always lower than men’s. In the economic-related posts, only 9% of the images are purely female, in contrast to pure male images, which account for 69%. Women have more opportunities to showcase in entertainment news images, but overall they are still lower than men.
You may be a womanThe scarcity of the figure is confusing, which is related to a larger social reality to a certain extent. For example, in the news reports about professional football teams, most of the images identified are male; in the US Senate and House of Representatives (Females account for 25%) In the report of , the identified female faces are of course much less than men.
Aside from the smaller details of these granularity, this study still reveals some alarming status: In Facebook news images, men are more prominent than women; in two or more people In the group image, men tend to be more than women. At the same time, men will occupy a larger visual space.
The researchers also measured the size of the female face and the male face in the image ( current technology can only capture the size of the face, ignoring the hair, The influence of factors such as jewelry and headwear). The results showed that the average face area of male faces was larger, and this difference caused the average facial size of males in the images to be 10% larger than that of females. In the image of Facebook, this shows that male characters can bring greater visual impact to readers.
Specifically, in economic-related posts, the average size of female faces is 19% smaller than that of men, but in entertainment-related content, the average size of female faces is 7% larger than that of men.
A machine vision tool like facial recognition is being used more and more widelyThe identification of gender in law, advertising and other fields is one of its basic functions.
In real life, recognizing the gender of the people around you is simple, but for a computer, what steps does it take to work?
How does the computer “see” your gender?
“After feeding the algorithm ‘thousands of image cases, as a ‘mature algorithm’, the facial recognition system itself can learn how to distinguish between men and women.” This answer can explain the above. Question, but for us outside the “black box”, it may not be easy to understand this learning process.
To better understand the rules in this process, the Pew Research Center conducted an interesting experiment in which they uploaded images of their center staff to the machine vision system and partially occlude the image content, hoping to Finding the law and finding out which facial areas will make the algorithm make or change the decision.
In this interactive challenge of “human-machine game”, you may also boldly guess which parts affect the judgment of the system?
First, enter a clear picture into the machine vision system. At this point, both the algorithm and you can clearly determine the gender of the person in the photo.
Next, there are a number of boxes in the photo that tell you, “Selecting a box means hiding the content in the image, your choice may affect gender judgment.”
Finally, when you complete your selection, the image will show all areas that affect the gender classification change.
Interested readers can visit the Pew Research Center website to complete this small experiment.
Portal: https://www.pewresearch.org/interactives/ How-does-a-computer-see-gender/
The following set of images is part of the results of an interactive experiment. When you select a purple or yellow area of the picture, it will bring about a decision change in the recognition system. In the current gender diversity, in real life, gender recognition is not easy, but Pew shows more clearly through this experiment, In the algorithm system, let the machine firmly say the testee’s Gender is too difficult.
After looking at this picture, what else can you find? — Sometimes, the part of the face that causes the model to change is perhaps very different from what we expected. For example, in the fourth picture, covering people’s faces will cause system recognition to change, but more often, the algorithm will produce the opposite interference area, which is actually the face edge, hair root, mouth angle and other areas.
From these experimental cases, you may have discovered that no uniform, stable law can explain this phenomenon. Sometimes, hiding the middle of a tested face can cause a change in gender recognition, but covering the other in the same way does not necessarily result in the same result.
Machine learning can greatly improve the efficiency of our data processing, but unlike traditional computer programs, machine learning follows a series of rigorous steps, and their decision-making methods are largely invisible and highly dependent. Used to train your own data. These characteristics may result in machine learning tools producing systematic deviations that are more difficult to understand and predict in advance.
From this perspective, the Pew Research Center used a simplified experiment to show how the data used to train the algorithm introduces hidden biases and unexpected errors into the system results. Researchers say that as algorithms are playing an increasingly important role in decision-making in human society, it is important to understand their limitations and biases.
What does “prejudice” bring?
Recently, 26 top AI researchers, including Turing Award winner Yoshua Bengio, asked Amazon to stop selling its artificial intelligence service Amazon Rekognition to the police in a public blog post. Anima Anandkumar, former chief scientist of Amazon’s cloud computing division, and others also joined the joint appeal.
Before, Deborah Raji, a researcher at the University of Toronto, and Joe Buolamwini, a researcher at the MIT Media Lab, wrote a research report that pointed out that Amazon’s Rekognition is better than judging skin color when detecting female genders with darker skin in images. The error rate of the lighter male gender is much higher. The research results have also been supported by scholars, but Amazon has written this for two people.Reports and research methods have raised objections.
Amazon Face Recognition System Test Accuracy for Different Skin Tones and Genders
Joy Buolamwini led an AI research project called Gender Shades. After researching the facial recognition systems of leading technology companies, it was found that all systems performed better on identifying male faces and all systems were identified. The accuracy of lighter faces is higher. The average recognition error rate for dark-skinned women is as high as 35%, for dark-skinned men, 12% for light-skinned women, and 7% for light-skinned women, and for light-skinned men, the error rate is less than 1%.
What might the “prejudice” of facial recognition systems bring?
Google recognizes this user’s friend as “Gorilla”
“Face recognition technology can be abused regardless of its correctness,” Joy said. Accurate or inaccurate use of facial recognition technology to analyze the identity, face, and gender of others may violate the freedom of others. For example, inaccurate identification may make innocent people obscured and unreasonably censored by law enforcement officials. This is not a hypothetical situation.
British non-profit organization “Big Brother Observing” (Big BroTher Watch UK) has published a report emphasizing that the face recognition technology used by the London Police Department has a gender recognition error rate of over 90%. Last summer, the British media reported such a news that a young black man was mistaken for a suspect because of a facial recognition technique and was searched by the police in full view.
A leaked report also shows that IBM provides law enforcement agencies with the ability to search for people in a video based on hair color, skin tone, and facial features. This news raises concerns that the police will use the technology to focus on specific races.
Ida B.Wells, a well-known African journalist and affirmative sportsman, is recognized as a male.
In order to reduce the time required to search for faces, law enforcement is using a large number of gender classifications. If the gender of the matching face is known, a simple dichotomy can greatly reduce the number of potential matches that need to be processed. Gender classification is being widely applied to police activities.
When these biased identification systems are widely used in social life, they can lead to even worse consequences.
Joy Buolamwini presents a speech entitled How I’m fighting bias in algorithms on TED
In the TED talk, JoY shared a little story with everyone:
Under the same lighting conditions, the facial recognition system can only detect participants with light skin color; only dark masks can be detected by wearing a white mask. “Before the artificial intelligence tool determines the identity of the face or identifies the expression information, the most basic premise is that the face is detected. However, the facial recognition system fails repeatedly in detecting the black skin individual. I can only comfort myself, the algorithm is not The racist, his face is too dark,” Joy said.
Where does the deviation come from?
If you compare the accuracy of the developer’s own claims with the research findings of the researchers, you will find an interesting thing: the data published by the company and the external accuracy of independent third parties are always different. So, what caused this difference?
Joy reminds us to focus on the bias of the benchmark dataset. “When we discuss the accuracy of facial analysis techniques, it is done through a series of image or video tests. These image data form a benchmark, but not all benchmarks are equal.”
The relevant person in charge of Amazon said that the company used more than 1 million face data as a benchmark to test the accuracy of the product. But don’t be confused by this seemingly large sample. “Because we don’t know the detailed demographic data of the benchmark data. Without this information, we can’t judge whether the choice of benchmarks may have the possibility of prejudice such as race, gender or skin color.”
Different systems have different recognition data for dark-skinned actors
Facebook has announced that its face recognition system is 97% accurate in dataset tests called Labeled Faces in the Wild. But when the researchers looked at the so-called gold standard dataset, they found that there were nearly 77% of men in the data set, and more than 80% were white.
In order to eliminate bias as much as possible at the data layer, Joy suggested that a more inclusive benchmark data set should be built. To balance the benchmark data, she lists the ten countries in the world with the highest proportion of women in parliament, with Rwanda leading the world by more than 60% of women. Taking into account the typical representation of the Nordic countries and a few African countries, Joy selected three African countries and three Nordic countries to balance the types of skin in the data set by selecting individual, dark-skinned individual data from these countries.
Based on this more balanced data set, they re-evaluated facial recognition systems from companies such as Amazon, Kairos, IBM, and Face++. In the August 2018 study, they found that Amazon and Kairos performed well in white male identification, but Amazon’s accuracy in identifying female faces of colored people was very low, at 68.6%.
Amazon’s facial recognition system puts a male label on this image of Oprah Winfrey and gives data confidence
Joy said that face recognition in the real world is more complicated and difficult than experimental detection. The benchmark data sets they build are not fully tested. “But it is like running games, in benchmarking. Excellent performance, at least you can guarantee that you will not fall off when you start.”
Even with the same benchmark, the accuracy number of the facial recognition system may change. Artificial intelligence is not perfect. In this case, it is a useful practice to provide users with more specific judgment information by providing confidence.
Face recognition technology has been widely used in large-scale surveillance, artificial intelligence weaponization and more law enforcement environments. However, this powerful technology is rapidly evolving without adequate supervision.
To reduce the abuse of facial recognition technology, Algorithmic Justice Alliance (Center on Privacy & Technology) launched “Safe Face Pledge” (Safe Face Pledge) Activities.
At present, many technology companies, including Amazon, have not yet joined this commitment. “According to our research, it would be irresponsible to sell facial recognition systems to law enforcement or government agencies.” Joy, one of the founders of the Algorithmic Justice Alliance, hopes that more institutions will join the “safe face” in the future. Commitment, can act responsibly and morally for the development of facial analysis technology.
After all, behind the algorithmic bias is actually our own human prejudice.
Reference link:
1.https://www.journalism.org/2019/05/23/men-appear-twice-as-often-as-women-in-news-photos -on-facebook/
2.https://www.pewresearch.org/interactives/how-does-a-computer-see-gender/
3.https://medium.com/@Joy.Buolamwini/response-raТ-and-gender-bias-in-amazon-rekognition-commercial-ai-system-for-analyzing-faces-a289222eeced
This article is from the public number:All Media (ID:quanmeipai).