News that Amazon’s face recognition tool mistakenly matched 28 U.S. lawmakers to criminals has raised concerns. Why did the face recognition tool make such an error? In fact, the machine and human, see strange foreigners will also have “face blindness”, can only identify “localpeople”, for “foreigners” from other countries, the recognition accuracy is low, how to solve this problem?
Citing publicly available personal information to conduct research and verify Deng Weihong team’s supply map
The research team is discussing The Images of Deng Weihong’s Team
The key is to make people aware of the faces of people around the world as much as possible, a process that is not easy to achieve. Deng Weihong, a professor at Beijing University of Posts and Telecommunications, told The China Science Daily that the more training data used and complete, the higher the accuracy of the face recognition tool. However, it is increasingly difficult to collect personal information from different countries. The absence of training data means that the face recognition tool only knows “acquaintances” and is less accurate for the people missing from the training data.
Recently, a new study by Deng Weihong’s research team has made new progress, the team revealed the prevalence of cross-country recognition bias in the current face recognition algorithm, constructed the face data set RFW to evaluate the degree of deviation, and proposed information to minimize the recognition deviation to maximize adaptive neural network to improve the recognition ability of the target domain. The findings were presented at the International Conference on Computer Vision (ICCV), hosted by IEEE, on 27 October.
The geographical difficulties of face recognition tools
Convolution neural network is one of the representative algorithms of artificial intelligence, and has a strong ability of image characterization learning. In 2012, the emergence of deep convolutional neural network in the field of computer vision, its emergence has greatly promoted the development of face recognition, and become the mainstream technology in the field of face recognition.
At present, most of the global facial recognition tools are based on the deep convolution neural network technology development, but the source area of the technology’s face data based on the facial features of Westerners, in the face of different target areas, that is, the facial information recognition needs of residents of different countries are often “inadequate”.
Mr Deng said research in the area had been slow for a long time because of a lack of benchmarking libraries. A face recognition tool, even with high local recognition rates, is difficult to pinpoint to humans around the world. This results in a strong geographical nature of face recognition tools.
To advance the study, Deng’s research team built a new test library, RFW, to scientifically and objectively measure deviations in face recognition.
Based on the RFW database, the researchers validated four of the most advanced algorithms in Microsoft, Amazon, Baidu, the business API and academia.
“This recognition bias does exist, and in some areas the error rate is even twice as high as in western countries. Deng Weihong said.
The paper’s review experts say RFW is more evenly distributed than the existing database, which would be a better benchmark ingress rating for cross-country identification.
To see if this deviation was caused by an imbalance in the distribution of training data, the researchers collected a training database covering human information in various regions of the world and found that the occurrence of the bias was influenced by both data and algorithms.
Wang Mei, the first author of the paper and a doctoral student at Beijing University of Posts and Telecommunications, explained that the database training data balance and algorithm are the same, but in some countries human facial information recognition is more difficult, resulting in lower recognition accuracy.
How can I improve recognition in situations where there is less data and facial recognition is difficult? The researchers didn’t give up, and they decided to further study the algorithm, using the algorithm to make people face recognition tools to make a case against three.
Learning by “conscious”
Traditional machine learning databases require manual labeling of personal information, which is at risk of privacy disclosure. The unsupervised domain adaptive approach for object recognition inspired the researchers.
This method uses unsupervised learning to map the source and target domains to the unaltered feature space of the domain, and improves the performance of the target domain. The researchers wanted to solve the problem by algorithms and let the machine sit on its own.
“It’s the equivalent of researchers putting the papers out and the machines answering them. Deng Weihong said.
Implementation is not simple, in the specific operational level, object recognition is different from face recognition. The source and target domains of object recognition can overlap, and information acquisition is relatively inexpensive, and sufficient source domain data enables identification tools to distinguish and distinguish the target domain information.
Therefore, the researchers proposed an information-maximizing adaptive network. Wang Mei introduced that, on the one hand, the method reduces the difference between the global distribution of the source domain and the target domain, on the other hand, it can learn the characteristics of the distinct target domain.
“That is to say, convolutional neural networks can ‘consciously’ learn the characteristics of the target domain face without supervision. Wang Mei said.
In order to solve the problem that the categories between the two domains do not overlap, the information maximization adaptive network uses the spectral clustering algorithm to generate “pseudo-label” and pre-adapts the network with pseudo-labelunder, which improves the performance of the target domain.
This clustering scheme is fundamentally different from other domain adaptive methods that do not apply to face recognition. Wang Mei explained that the new method can learn independently in the new target domain, without the need for human intervention, to avoid the risk of privacy disclosure.
To further improve the identification of network output, the researchers also proposed a new adaptive approach based on mutual information that creates greater spacing between the characteristics of the target domain in an unsupervised manner.
Unlike general supervised losses and supervised mutual information, the method has an unsupervised nature that allows it to utilize all unlabeled target domain data, whether or not the data has been successfully assigned to pseudo-labels.
Validate based on public data
Can this scheme improve the recognition of facial information by facial recognition tools to residents of different countries without supervision?
The researchers used public data from celebrities around the world for validation. The results show that the information-maximizing adaptive network can successfully apply recognition ability from the source domain to the target domain population in other countries, and the recognition performance is better than other domain adaptive methods. The study of ablation experiments found that the loss of mutual information plays an important role in reducing the recognition bias.
Wang Mei added that the information-maximizing adaptive network also has good generalization performance in cross-attitude and cross-scene applications.
In view of the better experimental results, the research team has released the RFW dataset to further advance the study.
At present, harvard University, Imperial College of Technology, Tsinghua University, Cisco, Huawei, NEC, IBM and other more than 20 countries of research institutes, enterprises, research teams have applied to use RFW for multi-human facial recognition research.
It is worth mentioning that the method still needs to train the model by using the label data collected by the source domain region while carrying out adaptive learning. This means that the source domain data is at risk of privacy disclosure. How to carry out adaptive learning of the target domain without transmitting the source domain data is a very worthy of study.
Deng Weihong said that the next step is to put forward a new algorithm with more generalization ability without collecting target domain data at all, which will directly improve the accuracy of the face recognition tool in the unknown target domain.
Related paper information: http://whdeng.cn/RFW/index.html