At the beginning of the new year, the confrontation between humans and artificial intelligence has already ushered in a climax. First, Google’s DeepMind’s Master’s online game in the Go game has won 60 games, which proves once again that the machine has surpassed humans in the sport of Go. A few days later, the "Best Brain" in the first quarter of the fourth season ushered in a heavy man-machine battle, the small robot implanted in Baidu brain against the world memory master Wang Feng. The theme of this competition is the face recognition of the New Year. The human player recognizes two photos, and the small ones answer all three photos, even including a twin photo, thus defeating the human master. The small and stunning performance is inseparable from Baidu's powerful artificial intelligence technology. At present, Baidu brain has the world's largest neural network, trillion-level parameters, hundreds of billions of training data and billion-level features, these powerful and powerful artificial intelligence systems, coupled with GPU parallel computing rich computing resources. Increased efficiency, so that artificial intelligence technology can really exert its power. The theme of this competition is face recognition, which is also an important area in current image recognition. It is mainly divided into four steps: face detection, face image preprocessing, face image feature extraction and face image matching and recognition. In the game, human player Wang Feng said that he also looks for and remembers some features when recognizing faces. For machines, the deep learning method is to learn different features layer by layer, from low-level features to advanced features. Whether it is Google, Microsoft, Facebook, Baidu and other technology giants, or many companies, they have actively explored the research and application of face recognition technology. Lin Yuanqing, director of Baidu IDL, said that in 2014 and 2015, FDDB Baidu was the first. At the end of 2015, it was better to have an error rate of 8% on the internally difficult test set. After the improvement in 2016, the error rate was reduced to 2.3. %. Finally, I hope to reduce the error rate to less than 1%. Baidu's face recognition system is a two-step training. The first step is a universal face recognition system. It is not specifically aimed at age. The most important part of this success is training a very powerful face recognition system. Our data is two million people, each with a hundred photos. We train our face recognition system with a very large amount of data. This is the data that Baidu has accumulated over the years. The iteration of this process is data and The algorithm iterates together. The very difficult thing here is that you have a very good algorithm to make the most of it. With this data, you can design very good algorithms to match these data. The identification of the New Year's Eve in the game is not a simple match, but a logical reasoning ability. Lin Yuanqing said: "It is very important to divide the face into 7 parts and play 72 points on the face. It is very important to learn which parts of the features. Collect similar data and tell the machine that this person grows up when he is a child. Let the machine learn what is an important feature." In terms of specific technology, Baidu IDL's face team uses end-to-end metric learning, that is, by learning a nonlinear projection function, the image space is projected into the feature space. In this 128-bit feature space, the distance between two faces of the same person across ages is smaller than the distance between two faces of different ages of different people. At the same time, taking into account the scarcity of cross-age faces, use a model trained with large-scale face data as a base, and then update him with cross-age data. This is not easy to overfit. Another very important factor in face recognition technology is feature point positioning. For this problem, Lin Yuanqing said in an interview with the heart of the machine: "One aspect is of course that we have to find a way to make very good feature point positioning. The algorithm, then to train a very good model. In some extreme cases, the algorithm may not be very accurate positioning, but we hope that the latter identification module has a certain fault tolerance, then our approach is during training Make some data, artificially generate some errors in the positioning, and then put the data in the deep learning model to train, so that the last trained model has a certain fault tolerance for the positioning error." In addition, the game was played on the spot, and the picture seen by the small one was exactly the same as that of the human player, not the access of the image signal. The complexity of TV shows increases the difficulty of face recognition. For example, angles, expressions, live light, shadow makeup, and accessories, but the small performance is still accurate. Wei Kunlin, the strongest brain judge, commented on Baidu's face recognition technology: "Baidu's powerful point on the stage of the strongest brain is real-time comparison. Everyone has not seen the material of the challenging projects on the stage. In advance, the two sides only know The approximate project of the challenge is left to Baidu engineers to train artificial intelligence based on data from normal people and ordinary scenes." The small performance in face recognition makes us look forward to its future applications, and Baidu's advancement in face recognition products and applications has begun. At present, the face recognition technology based on face recognition has already landed in Baidu Building; banks will also use this technology for remote identity authentication; Baidu's face recognition is also unveiled at the Wuzhen Theater Festival. For this game, Lin Yuanqing did not like to use "defeating" to describe. "Whether we win or lose, we are going back to continue research." He said, "In the next five years, ten years, even twenty or fifty years, we will coexist with artificial intelligence technology. I hope we can use these technologies well. Helping humans solve problems instead of making these technologies the opposite of humanity." Li Yanhong also said: "Whether you win or lose, you will make a breakthrough contribution to the technological development of artificial intelligence." In fact, for every man-machine competition, there will always be some non-professional interpretations of threat theory or fatalism. This is not only irresponsible to human beings, but also affects the normal progress of artificial intelligence technology. Just as the CNN applied to image recognition borrows somehow from the principles of human vision, this competition will also enable Baidu's technical team to get enough inspiration and clues from human players to promote the advancement of artificial intelligence. . The game only proves that the machine is on the task of face recognition. Based on the most advanced technology, it can do better than humans, bringing us more imaginable space for application. More importantly, artificial intelligence is brought right in front of everyone and prompts us to expand the boundaries of intelligence. FME Connector,RF Connector,PCB Mount,Angled/90° Dongguan Yiyou Electronic Technology Co., Ltd. , https://www.yiucn.com