Editor's Note: A year ago, Facebook released a photo sharing application, Moments, which recently closed the iOS version of Facebook photo synchronization and pushed the Moments application. This application uses face recognition technology. However, Yann Lecun, head of the Facebook artificial intelligence laboratory, introduced the principles of Moments to us in an easy-to-understand way. In addition to simple face recognition technology, Facebook will use more advanced computer vision technology and AI technology to provide users with more Convenience, such as trying to develop computer empathy capabilities, of course, these convenient applications behind the need for strong algorithms and tedious training process support. Let us look forward to the future that computers can better understand humanity and make our work and life more convenient.
Facebook's strong photo sharing application, Moments, uses image recognition technology to enable users to scan private photos and create private group albums. When six people took pictures at the event site, Moments made it easier for all participants to share photos with each other, eliminating the cumbersome process of sending snapshots to everyone by mail and creating albums. Of course, this kind of application is not like finding a cure for cancer. It is called a feat, but there is an impressive technology behind this convenient photo sharing application. Facebook took a few years to develop it. Out of this technology.
The key point of the Moments application is to have Facebook's algorithm as a support, so that this application can identify the same person's face in different photos , so that Moments can know who is at the event site. This requires expertise in computer vision. Google, Microsoft, Baidu, and other companies have been working on the field of computer vision for a long time. Applications range from self-driving vehicles to point-and-click web products like Microsoft's "How old I look". Applications, changes ranging.
The release of Moments represents that Facebook is sharing its success in the field of computer vision research with the world. Facebook face recognition accuracy can reach 98%, the recognition speed is also extraordinary, can identify your photos from 800 million photos in less than 5 seconds. Finally, even if you can't get your frontal shot (or your face doesn't appear in the photo at all), Momemets can easily identify your photo, thanks to a machine learning algorithm that takes care of both Other elements related to photo data.
Moments' source of development
Fortune once interviewed Yann LeCun, head of Facebook's artificial intelligence lab. During the interview, he learned how his team helped a computer understand "who you are" and what will be the next step in Facebook's AI research. When LeCun talks about computer vision, perhaps we must first understand that although the process of church software recognizing objects is similar to people's looking at external things, we still cannot simply equate computer vision with people looking at external things.
For example, Facebook's face recognition technology doesn't recognize you. Only when the two people in the photo are the same person can this technology identify the identified object. Face recognition technology is actually a completely separate step.
Since Facebook is mainly for the establishment of a close relationship between people, its computer vision technology focuses on the recognition of human faces and is not used to identify cats, vehicles, or other physical objects. In order to achieve this goal, Facebook uses a database called "Labeled Faces in the Wild," which consists of pictures of celebrities and politicians. The database contains 13,000 photos in which people use different hairstyles, different clothes, sometimes wearing glasses or other decorative objects. Facebook uses this image set to train its machine learning algorithms. Other companies have also used this dataset, and some universities use this image collection training system to achieve an accuracy of 98%.
So how exactly did Facebook pass a photo of Angelina Jolie to a machine, and then, with the help of this photo, make this machine identify your sister from different albums? LeCun can answer this difficult question for us. About 20 years ago, when LeCun worked in the Bell Labs, he accidentally thought about how to teach computers to "see" the world. This idea was not applied outside of academia until 4 years ago.
How to teach computers to "see" the world
The technique used in computer vision is called a convolutional neural network . The designation is derived from a mathematical operation called "cyclotron" and inspired by the learning principles of the human brain . The human brain learns by establishing connections between neurons. The higher the frequency with which a signal is transmitted to a neuron, the tighter the connection between neurons. Similarly, when the computer establishes a similarity link for two images, it will assign weights to these similarities. In a convolutional neural network, the goal is to train the machine so that it can recognize changes in weights between established connections so that the computer can more accurately determine whether there is a match between images and images.
This training identification process is quite complex and involves different calculation methods. Using these calculation methods can determine which features in an image are important for identifying image information. For example, if you want to train a computer to recognize faces, the pixels of the image background appear less important. It is incredible that the machine can learn on its own, identify which features of the image are the most relevant, and then summarize these important features. However, it is still necessary to artificially promote the computer to identify reasonable ways to assign weights to similarity. Once the modeling is successful, the computer will have the ability to summarize important features.
It will take several days to complete this training process on a well-configured computer.
When Confucius University Professor Geoffery Hinton led his research team to use convolutional neural networks to win the competition in image recognition algorithms, convolutional neural networks became almost the cornerstone of all computer vision research. Later, Hinton’s research team and its new company were acquired by Google. Hinton won the game with a test error of 15.3%, and the second winner was 26.2%.
Looking forward to computer empathy
Facebook's automatic face recognition technology helps users protect their privacy from infringement. For example, when automatic face recognition technology is widely used, you will be notified every time you upload a photo to Facebook. For example, if you inadvertently appear in the background of a photo taken by a tourist in Times Square, you can be notified in a timely manner, and you have the option to blur your face in the photo and make it unrecognizable. For kids, if this happens, Facebook will automatically turn on blurring or removal. Lecun said that Facebook is very interested in such application tools, but he also stressed that Facebook's interest in the field of machine learning far exceeds image recognition technology.
The goal of Facebook is to make computers empathetic. Obviously, computers cannot perceive people's behavior, but they can train computers to have the ability to recognize human emotions and human reactions. When the computer's ability to understand reaches this level, when you are about to upload your drunk photos, Facebook will send out a prompt to confirm if you really want to do so.
"Computer empathy far exceeds face recognition technology," LeCun said. “We don’t care who’s in the picture. We will use other types of image recognition technology to train these technologies through different means so that the computer can recognize it. This picture looks embarrassing and prompts you whether you really want to Publicly available to the public."
Of course, Facebook does not have the ability to develop this kind of image recognition technology at the present stage, but LeCun put forward these concepts as experimental ideas, indicating the future direction of Facebook artificial intelligence research. Of course, this kind of image recognition technology that can be achieved with only one algorithm will really bring deep anxiety to people. Nowadays, because of people’s concern for personal privacy, Facebook has not yet promoted such auto-tagging applications in countries such as Canada and the European Union. To make a computer able to guess your intent to share photos in a matter of seconds, or to make the software try to analyze your jokes, and to understand the jokes in your jokes, you also need to consider a potential change influencing factors.
“Our goal is to make machines more intelligent and understand text, images, video and email. In a digital world, we want machines to understand what happens,†LeCun said. Because so many data contents can be received every day, people are inevitably overwhelmed. The efforts of LeCun's research team will help people obtain content that is closely related to their interests. To achieve this simple goal may involve a complicated solution: Make sure you see the information you need on Facebook .
“Making the machine understand humans is a major task that Facebook has been trying to accomplish,†LeCun said.
This article is compiled by Lei Feng Network (search "Lei Feng Network" public number) exclusive compilation, refused to reprint without permission! ·
Easy Electronic Technology Co.,Ltd , https://www.yxpcelectronicgroups.com