In the comments section under Andrew Ng Weibo, someone asked if there were any "shady" comments pushed to the front row.
In view of several points in the black post, Lei Feng. Com to have a look. What are the chances of this shady? Black posts are mainly aimed at the following points:
1. The rule has been temporarily changed.
The night before the program was recorded, the difficulty of the program suddenly increased, the video brightness was lowered by 20%, and the difficulty of the challenge increased several times, which put Wang Yuheng at a disadvantage.
Wei Kunlin also explained at the competition site: "Wang Yuheng asked the program group to reduce the brightness and make the photos more blurred by 20%, but this difficulty (for people) has increased too much. Maybe the machine will not be affected by brightness in this respect."
Regarding the temporary change of the rules, Wang Yuheng expressed his thoughts in the live broadcast: "To increase the difficulty, what I want is whether the darker the computer, the louder the noise. If I lower the brightness, will it come back? Just like if you give a person beauty, it may be difficult to change this photo back to the original picture. "
Wang Yuheng's motivation is actually to make the robot more difficult. He won three times in his previous rehearsal. In this case, his motivation to increase the difficulty is understandable. He doesn't want to draw, he wants to win, and he wants to make the computer go wrong. But manually adjusting the brightness of photos may have no effect on the computer at all.
Wei Kunlin said: "Artificial intelligence uses contrast to solve the problem of recognizing contours. Even if the whole thing gets dark, the contrast is still the same. Only this thing can get the outline of the line out. "
A doctoral student at eecs told Lei Feng.
"If the values of all pixels are reduced proportionally by manually adjusting the brightness, then compared with people, the impact of manual adjustment on the computer is not too great. But if it is because of the change of shooting environment, such as the sky getting dark, then artificial intelligence will be affected. Professor Wei mentioned that it is correct to recognize faces by edge contour. Like some traditional face recognition algorithms, feature points are detected by relative brightness, so reducing brightness has little effect on such algorithms. However, for deep learning, because we still don't fully understand what features it extracts for recognition, and most of the open source training photos have moderate brightness, we do find that the brightness of the shooting environment is influential in the evaluation of some deep learning algorithms. "
In other words, if a small team of engineers uses algorithms to restore the brightness of photos, then the difficulties faced by humans and computers are actually unequal. In addition, the doctor also said: "It is only fair to dim the nearby lights when recording videos."
But the video was finished a month ago. Wang Yuheng said in the live broadcast: "I mistakenly thought that ears and sideburns are the only places that are not covered, so they can best reflect the real information. I didn't expect this (video) to be a month ago. I also cut my hair in the middle and got acne and acne. " The rules were changed the night before recording, so it was impossible to fundamentally adjust the brightness of the video at that time. From this point of view, Wang Yuheng mistakenly thought that changing the rules would increase the difficulty of machine recognition without knowing much about the operating principle of machine learning, but in fact, he didn't. In his own words, he is "smarter than smart"
Wang Yuheng changed the correct answer to the wrong one.
In the video, we see that Wang Yuheng has made many choices between the two options. But he often erases the correct answer and finally chooses the wrong answer. In the end, all three of his answers were wrong. In view of Wang Yuheng's previous performance, it is unusual to be wrong three times, so some people suspect that Wang Yuheng was "forced" to erase the answer and even "restored" the angry picture of Wang Yuheng at that time.
Wang Yuheng himself explained in the live broadcast: "If you let me write only once, maybe I am right. What I erased was all the answers I had written before, and the rest was what I calibrated. " Wang Yuheng used human ears and sideburns as "calibrators". This information will change with time, which is the main reason why he was wrong three times.
3. The small dialogue question answering system was questioned.
In addition to the competition, Xiao Di also interacted with the host and guests at the scene, showing the humorous and lovely side of the robot. However, it seems that this question and answer system is not automatically generated by the computer.
First of all, we don't see Xiao Di using "wake-up words" to start the dialogue during the interaction. It seems that people can talk to Xiao Di without saying "wake-up words", which is impossible in the current NLP field. One of the details is that when the host asked Wang Feng if he was confident in his answer in the first program, he turned to Xiao Du and asked him, "Have confidence in your answer ....." Before the words were finished, Xiao Du answered "Yes" by himself. The host couldn't help but vomit: "I haven't asked him yet, so he said."
A doctoral student engaged in NLP research told Lei Feng. Com: "At present, it is impossible for the machine to generate such an answer. Most people don't join the answering function because they are not sure when to answer. So I guess there is a desktop and staff control in the background. "
4, the second round of small degree is "outrageous".
In the second round, Du Fu and Wang Yuheng got their answers wrong. The correct answer is No.8 in the middle, Wang Yuheng chose No.7, and Xiao Du chose No.29. It can be seen that No.7 is similar to No.8, but the short answer No.29 is far from No.8.
For this, Baidu engineers seem to be surprised. On the Zhihu, Weng Renliang, who claimed to work for Baidu IDL Face Group, answered. He said that he had participated in the super brain project team and described the team's response at that time as stupid.
Why is there such a big error in the system? Lei Feng. Com consulted several experts in the industry and did not get a clear conclusion. Lei Feng. Com also further contacted the engineers of Baidu IDL, and has not received a reply so far.
Why was Xiaodu hacked this time?
It is also a "man-machine war" similar to PR. Compared with the game between AlphaGo and Li Shizhen, why Google has not been questioned, but has achieved the effect of shocking the world (because of the influence of AlphaGo, 20 16 even became the "first year of artificial intelligence in China"), Baidu has an authoritative figure in the field of artificial intelligence, Andrew Ng, and even the party Wang Yuheng has a "clarification live broadcast".
Why does Baidu's "man-machine war" attract so many questions?
One of the most important parts is that the "broadcast form" has discounted the authenticity of the game. The match between AlphaGo and Li Shizhen was broadcast live. From the perspective of a third party, the audience can see the whole game, and the details such as every move and thinking time of both sides are presented in an unobstructed view. Therefore, the dramatic scene of Li Shizhen's One Hand of God has been talked about by people until now, and people will not question whether it has been arranged because the audience are all "witnesses".
Super Brain is a recorded program. In order to create a dramatic plot, the program group will edit the game later and add lighting and music special effects. Therefore, for the audience, the authenticity of the competition process will be discounted, which is essentially determined by the "broadcast mode".
In addition, in the first program, Wang Feng was temporarily ordered to play. It is unfair for human beings to compete with a well-trained machine without preparation. Another point is that the authenticity of a small voice question answering system is not high, but the program creates a small effect of chatting with the host and guests. Ordinary viewers may really think that the ability of AI has reached such a high level, which will make people have unrealistic expectations for AI, which is unfavorable to the development of the whole AI.
Lin Yuanqing, director of Baidu's deep learning laboratory, said in an interview: "In the past few years, we have invested heavily in Baidu's brain. The original intention of participating in the super brain this time is actually to see to what extent our technical level has reached. "
Super Brain will always be an entertainment variety. If Baidu wants to be a successful PR like AlphaGo, it can actually consider holding the control of the competition in its own hands, inviting a group of human experts to participate in the competition organized by itself, doing a good job in peer review of the competition system and making a thorough contest. Maybe it will be much better.
Andrew Ng said in an interview: "Although Xiao Di can do very well in face recognition, AI can't do what Wang Hao and any human can do now." However, it seems that such a message of "Please treat AI rationally" cannot be conveyed to the public through entertainment programs. On the contrary, it will only make AI's enthusiasm bigger and bigger.
Generally speaking, this small-scale participation in "Super Brain" was hacked because this program "some lies were wrapped in the truth", which is really a bit difficult to convince the public. Of course, if you regard it as an entertainment program, it's nothing.