The robot once again fights against the strongest human brain. This time, adult face recognition

On January 6th, Jiangsu Satellite TV's "The Strongest Brain" staged a wonderful man-machine showdown. This battlefield is no longer a Go, but a face recognition.

The human representative of the battle is Wang Feng (microblogging), who is the master of the memory of the post-90s world and the rotating chairman of the "Best Brain" Hall of Fame.

In 2015, he participated in the "Best Brain Season 2" as the captain. In the "Best Brain" Sino-German International Competition, Wang Feng led the Chinese team to win the German team 4:0, and he defeated the fast memory poker. World record.

One side of the machine is Baidu robot "small degree", Baidu brain has implanted many research results in the field of artificial intelligence.

"Baidu Brain" has built a super-large-scale neural network, with trillion-level parameters, 100 billion samples, and hundreds of billions of character training, which can simulate the working mechanism of the human brain. Baidu's brain now has an advanced development of IQ, and even surpasses humanity in some capabilities.

In the international evaluation of face recognition technology, Baidu can achieve an accuracy of 99.77%, and in 2015 it has won two world firsts. The first game of man-machine battle is PK face recognition.

"Small" will compete with Hall of Fame players for three games, mainly in face recognition, voice recognition above PK, the first three man-machine wars, using three wins and two wins, if Baidu brain wins, will participate in the final competition The brain king is hegemony.

Round 1: Identification across ages

The guest (Zhang Ziyi) selected three difficult photos from the 20 bee girl team's childhood photos, and the player matched the selected childhood photos with the existing adult girls through the dynamic video performance. If you choose the correct one, you will get 1 point.

The bee girl team has a large number of people and each person has a make-up performance on the court. It does not rule out the interference caused by micro-shaping, Dai Meizhen and other factors.

In addition, the selected childhood photos are in the range of 0-4 years old, and the age of the current adult girls team is relatively large.

At the same time, real-time photo transmission, live camera capture of face image shaking, lighting interference and other factors will affect the recognition accuracy of artificial intelligence.

The most difficult thing is that there are twins in the Bee Girl team, which happened to be drawn by the guests.

In the end, Wang Feng, who was not informed beforehand, failed to distinguish the differences from the twins, resulting in a wrong judgment. The first round scored 0 points.

The Baidu robot gives two results. The difference is that the similarity is only 0.01%, and the higher similarity is finally proved to be the correct answer, thus getting 1 point in the first round.

After the first round, the score of the man-machine battle was 1:0, and human beings were temporarily behind.

The second round: thousands of faces across age recognition

The man and the machine jointly observed a viewer over the age of 30, and then found him from 30 elementary school photos. This round added difficulty on the basis of the previous round, so the score increased, and the correct one scored 2 points.

The sample size of this round is large, and 30 group photos need to find corresponding people in 1000-2000 faces, and the age span is also covered in the age groups of 80 and 90.

In the end, the machine and Wang Feng correctly identified the audience selected by the guests in the photo, and both scored 2 points. With the first round of scoring, the machine eventually scored 3 points and Wang Feng scored 2 points.

After two rounds of competition, Baidu robots won with a slight advantage, and Wang Feng paid a price for the difference of one in ten thousand twins.

Technical difficulties in face recognition

The human brain has the ability to recognize faces since millions of years ago, and the machine has no intuition, and it does not have a long history of evolution. It can only be learned by analyzing data.

The computer only knows 0 and 1, so it must find the law of human intuition through countless times of learning and turn it into 0 and 1 stored in the brain, thus simulating the process of human thinking through intuition.

The difficulty of face recognition technology research is different from ordinary image recognition. In terms of human facial features, each person's facial structure is similar, which is disadvantageous for using human faces to distinguish human individuals, and there are some special circumstances, such as twins and even multiple births.

The second is the external factors such as expression, lighting conditions, and plastic surgery. Different expressions, angle observations, effects of lighting conditions, face coverings, such as masks, sunglasses, hair, beards, and even cosmetic, P-pictures, etc., increase the difficulty of face recognition.

The identification of twins is technically more difficult.

Face recognition is to take as many points as possible on the face bone, and compare these points with the face that they have stored by the computer, and the difference is judged. Because the twin bones are too similar, the difference is particularly subtle, so if you don't have enough facial bones, you can't identify them.

Main steps of face recognition

(Taking the competition as an example, the flow chart of the small-scale identification of bee girl members on the spot)

The specific breakdown is as follows:

Step 1 Face Detection:

According to the characteristics of the eyes, eyebrows, mouth, nose and other organs and the geometric positional relationship between them, the face is detected, that is, whether a person's face is judged in an image or a sequence of images (such as video), and if so, the person is returned. Information such as the size and position of the face.

Step 2 Face image preprocessing:

The original image acquired by the system is often not directly used due to various conditions and random interference. It must be pre-processed with grayscale correction and noise filtering in the early stage of image processing.

The preprocessing of face images mainly includes face alignment, face image enhancement, and normalization.

Face alignment is to obtain a face image with a correct face position;

Image enhancement is to improve the quality of face images, not only to visually clear images, but also to make images more conducive to computer processing and recognition.

The goal of the normalization work is to obtain standardized face images of the same size and the same range of gray values.

[Preprocessing of Face Image]

Step 3 Face image feature extraction:

Face feature extraction is performed on certain features of the face. Face feature extraction, also known as face representation, is a process of character modeling a face.

Step 4 Face image matching and recognition:

Face recognition is to compare the face feature to be recognized with the obtained face feature template, and judge the identity information of the face according to the degree of similarity. This process is divided into two categories:

One type is face recognition, which is a one-to-one process of image comparison. One-to-one comparison is performed between a person's face and a designated person's face, according to the degree of similarity (generally whether or not a certain amount of quantization is reached or exceeded) The credibility indicator/threshold is based on whether the two are the same person.

The other type is face recognition, which is a one-to-many process of image matching and comparison. Compare someone's face with the faces of multiple people in the database (sometimes called "one-to-many" comparisons), and identify the person's identity based on the results of the comparison, or find the most similar face. And output the search results according to the degree of similarity.

Baidu brain enhances the method of cross-age face recognition

There are many factors affecting face recognition, among which the factors affecting face detection are: illumination, face gesture, and occlusion;

Factors influencing feature extraction are: illumination, expression, occlusion, age, and blur are the key factors affecting face recognition accuracy. There are more influencing factors in cross-age face detection.

In general, in face recognition across ages, intra-class changes are often greater than inter-class changes, which creates great difficulty in face recognition. At the same time, training data across ages is difficult to collect. Without sufficient data, deep learning-based neural networks are difficult to learn intra- and inter-class changes across ages.

Based on the first point, Baidu IDL's face team chooses to use the metrics learning method. That is, by learning a nonlinear projection function, the image space is projected into the feature space. In this feature space, the distance between two faces of the same person across ages will be smaller than the distance between two faces of different ages of different people.

For the second point, consider the scarcity of cross-age faces. Baidu used a model trained with large-scale face data as a base, and then updated him with cross-age data. This is not easy to overfit.

Combining these two points for end-to-end training can greatly improve the recognition rate across ages.

In addition, Baidu's face test set has 200 million pictures of 2 million people as training sample data.

Expert Reviews

Baidu Chief Scientist Wu Enda: Xiaodu not only represents Baidu artificial intelligence, but also represents China.

Baidu Chief Scientist Wu Enda

The world's top scientists can only understand part of the human brain's operating mechanism. Baidu's artificial intelligence algorithm has fewer human brains and more based on data analysis and deep learning.

In this competition, the competition we chose was very difficult for the machine, involving face recognition, speech recognition, etc., but in fact these are relatively easy for humans. People can make good judgments through intuition. For example, when you meet someone, you can recognize who he is without thinking. But the machine has to train from a lot of data, and some projects even need to identify unclear, old photos, so I think this is a huge challenge for the machine.

The face recognition skill, the human brain has been in existence since millions of years ago, and the machine has no intuition, and there is no long history of evolution. It can only be learned by analyzing data. So this skill is very difficult for even the most advanced AI technology in the world.

Today, based on powerful data analysis, it's easy to identify two recent photos. However, we don't have a lot of data to analyze for photos that identify facelifts, makeup, or a span of more than a decade. So this is a worldwide challenge for face recognition technology and one of the biggest challenges in today's game.

There are very few top players in the world of chess, and face recognition is available to everyone. This man-machine war is a top-level face recognition player and an artificial intelligence competition that is good at board games. It is fair.

Human beings are entering the era of artificial intelligence. In the near future, artificial intelligence technology can be applied to lost children, and powerful artificial intelligence creators are still human.

At the moment, you can't fully understand human thoughts, but you should learn from Wang Feng and the top brain of the Hall of Fame to better serve humanity. Smallness not only represents Baidu artificial intelligence, but also represents China. This man-machine battle was the first time Baidu’s brain appeared in a public game. The result was unclear and it was only waiting for it.

"The strongest brain" Dr. Wei: The human intelligence is behind the people, is the crystallization of the work of scientists

People think that the simplest thing is very difficult for artificial intelligence. For example, sports, although you will climb stairs when you are three years old, but now we do not know how to make the robot climb the stairs as smoothly as people, especially when the parameters of the stairs are unpredictable.

People can climb a variety of stairs, in different lighting conditions, different physical conditions, and so on. But the robot can't be as smooth as people now. Evolutionally, the movement, including movements like climbing stairs, the brain learned very early.

And people learn to play chess on the evolutionary brain very late. Therefore, for people, the stairs are a little easier, and the Go is a little harder. But it may be easier for the machine to play chess, and it is more difficult to get on the stairs.

Perception and movement, this is what humans are good at. We have been doing this for millions of years, and we are not good at the abstract thinking ability represented by logic and computing. The machine is not good at perception and movement. You will find that robots can play Go or write down a huge amount of information, but there is no way to move like a human, or to perceive this complex and rapidly changing world like a human being.

What artificial intelligence is currently good at is a rule that clearly defines what he can solve, that is, Go. Go is a rule, he has a target state, that is, I can take up more than you, I am surrounded by you, chess is even more, I will kill you. At present, many problems that artificial intelligence algorithms can solve are regular, or the target state is clearly defined. But in human society, there is no rule in what the human brain wants to achieve, and even accurate target states cannot be known in advance.

A lot of people's skills, that is, they will continue to improve. Except that some are physiologically aging and your muscle system is aging, there is no way. But many skills, if not restricted by physical body, many skills are better and better. In addition, the overall IQ of human beings is increasing year by year. The so-called Flynn effect, the average IQ is increased by about 3 points every 10 years. Of course, the main improvement is the ability of abstract thinking.

Artificial intelligence is also behind people, it is the crystallization of the work of many engineers and scientists. The machine wins humanity, which is the inevitable result of the development of science and technology. This day will come sooner or later, just coming early and late.

The development of science and technology is actually beyond our imagination. This day will come sooner or later, including general artificial intelligence that we are not currently able to achieve. Only the current engineers are doing a regional attack, and some hard bones are awkward. On this stage you can say that in some areas artificial intelligence has reached its peak.

Artificial intelligence surpasses humans in face recognition. It should be 2012, it is said that face recognition exceeds the human average and is a milestone. Now, Baidu’s brain is beyond the reach of a group of people. It can be said that in this professional direction, the accuracy of artificial intelligence has reached a high level, and the next step should be to improve the efficiency and energy consumption of computing.

When any new technology appeared, the people panicked, the car panicked, the train panicked, and the computer panicked. This is the ultimate panic, because the panic in the car is just that this thing is very fast, can kill me. The same is true of the train.

The first thing people think of is their own unemployment. The automated factory thinks about the unemployment of industrial workers. The emergence of artificial intelligence may make many general intellectual activities (including many white-collar jobs) and even professionals (including doctors in certain fields). The work is threatened. However, I feel that the overall unemployment rate of mankind will not necessarily decline. Some jobs are dead and new jobs are created.

Lin Yuanqing, director of the Baidu Deep Learning Institute: Defeating humans is not an end

In recent years, Baidu has invested considerable effort in artificial intelligence to do technology research and development. We want to compete with people in areas where people are good at it. In the end, what is our level, and whether it is close to people in these areas, or Said that there is a big gap.

Defeating humanity is not an end. I hope that we can evolve a good technology to serve humanity.

In recent years, Baidu has invested considerable effort in artificial intelligence to do technology research and development. We want to compete with people in areas where people are good at it. In the end, what is our level, and whether it is close to people in these areas, or Said that there is a big gap.

Industry Control Circuit Board

Industry control printed circuit boards (PCB)

1. For industrial control, special or general-purpose circuit board, the bottom circuit is completed, and IO is reserved. After buying the industrial control circuit board, the input and output ports reserved on the circuit board are connected to the user's own device, such as motor, solenoid valves, sensors, thus completing the function want to accomplish.
2. The special circuit board is a board specially designed for a certain function, such as the temperature control circuit board. After buying it, the input port is connected to the thermocouple, and the output port is connected with a heating contactor or a solid state relay to control the heating device to complete the temperature control. There are flow control boards, motion control boards, etc.
3. The universal control board, most of which can be programmed, after the user's own secondary development, completes a specific function, and has a wide range of uses. For example, plc is a general industrial control board. After the user writes the program and compiles the input, various functions can be completed. Digital input, analog input port, high-speed counter port, digital output port, analog output port, and some also have communication function.

Industry Control Pcb,Electronic Control Board,Industrial Controller Circuit Board,Industrial Control Printed Circuit Board

Chuangying Electronics Co.,Ltd , https://www.cwpcb.com