"DreamNet? What kind of network structure is this?" Li Feifei carefully searched through the descriptions submitted by each team.
It can be seen that almost all teams' submissions describe, "Our structure is based on AlexNet", "We imitate AlexNet", "We combine AlexNet with traditional support vector machines/../..".
The methods are similar, and the results are naturally very different.
What remains is a description of how to integrate several models together.
Turning to the team Dream, the personnel composition column only has two names: "Fanqi Meng, Huang Tang" and no specific organizational information.
Looking at the introduction of the team's algorithm, there are only a few words.
"We proposed a new DreamNet that enables networks as deep as hundreds or even thousands of layers to be trained. Due to the lack of sufficient equipment, we only obtained a single weight for submission."
Meng Fanqi did it on purpose. Of course he could have briefly introduced the residuals or other major tricks in the team's submission description. But why should we do this? We must first arouse the curiosity of these people in the industry in order to gain the greatest attention.
Classification top-5 error rate, 4.9%.
Such a simple number directly hit Li Feifei's heart.
The error rate of the algorithm is lower than 5.1% of humans. This has been her goal since she promoted this competition in 2009, and it is also a major reason why she organized this competition.
She had even made plans to stop organizing IMAGENET if a certain algorithm succeeded in breaking through this level.
But she never thought that this day would come so quickly.
Just 400 days ago, the best traditional methods could only achieve an error rate of about 25%.
Last year, AlexNet's error rate of about 16, she thought, was a leap forward. Such progress can be said to occur only once in decades.
She never expected that this year's results would give her an even more exaggerated surprise.
DreamNet, DreamNet that breaks through the human level.
Li Feifei felt that the name was really appropriate, and the result was truly dreamlike.
The test result mAP: 0.34673, which is a dimensionality reduction attack on the second place of 0.22581.
My God, who are these people? Li Feifei thought hard and couldn't think of any AI laboratory where he had heard these names.
"Why don't you briefly describe the content of the algorithm!" Li Feifei slammed her hand on the table twice. No one can understand and understand the value of this algorithm's performance better than her, a data collector.
From the moment she saw the result, her curiosity was like a wild horse that could no longer be controlled. I really want to know, I really want to know. Even if it's just a description or two that's specious, or even if it's just one or two specific terms that are unclear.
It's a pity that Meng Fanqi is well versed in media propaganda. Between the lines, "We have proposed a new DreamNet" means that this structure is completely different and has changed greatly compared to AlexNet, in order to enhance the curiosity of readers.
Then a bombshell was thrown, "We enable networks as deep as a hundred or even a thousand layers to be trained." Compared with last year's 8-layer champion, this number is really shocking, and it makes technical practitioners itchy. .
Finally, I pretend to be humble, but in fact I mean that we only used relatively limited resources to obtain a result, and we will not follow you in the multi-model fusion and integration.
There's so much difference in performance, it's really unnecessary.
"Put the results page online as soon as possible." Li Feifei replied to Deng Jia, thinking to herself that it was no wonder that Deng Jia, who was calm and down-to-earth, had just been so rash, "I cannot bear this crime alone, I will make you all anxious. Like ants on a hot pot."
The best brains are very curious. This curiosity about the unknown and the pursuit of better drives them to make results that can change human life one after another.
It may be easier to tolerate slightly more approximate results. Because I feel that if I think about it more carefully, try to test it, and add a few tricks, I might be able to achieve the same results.
And an even more exaggerated improvement than last year's AlexNet is in front of us. Even the most conceited person has to admit that he can't do it in a short time.
A larger gap makes people less suspicious and jealous, only more pure curiosity and admiration.
Meng Fanqi is not prepared to change the names of most technologies, but only calls the result of the residual network trained by mixing some other methods such as batch normalization, second-order optimizer, etc. DreamNet.
First, because he was reborn from a ten-year dream, returning to 2013 to submit the residual network was like a dream. In fact, he occasionally worried whether he would wake up from this dream at any time.
The second reason is because my surname is Meng, which is homophonic to Meng. If you really want to submit under the name MengNet, it would be too shameless, and it is absolutely unnecessary.
After the results of the competition came out, Tang Juan and Liu Xu also felt that life was like a dream.
Even if Meng Fanqi's updated code and tips can always improve the performance of the model or the training speed, even they have thought countless times, can they really win the championship?
But deep down, they had always just doubted it and never really believed it.
Not to mention winning the championship with such a crushing attitude.
"Microsoft, UC Berkeley, St. Petersburg, IBM, Tokyo University, National University of Singapore, Oxford, Toronto." Counting down the list, how many world-class universities and institutions are left far behind.
Even Li Kun, one of the three giants of AI, personally led the team to participate in the competition this time, but he could only fail.
At the same time, just as Li Feifei asked Deng Jia to publish the results page as soon as possible, three people from Google in Mountain View, Silicon Valley, near Stanford University, immediately paid attention to the results.
They are last year's champion, the Hinton team, which was acquired heavily by Google.
Alex is tall and thin, with short hair and a square face. Dressed in a classic programmer's suit, his words and deeds are crisp and clear, and he feels very intelligent and capable.
Hinton has three-dimensional facial features, a slightly hooked nose, and looks a third like Snape. When he is expressionless, he looks very majestic and intimidating, but when he smiles, he is very contagious.
But at this moment, he really couldn't laugh.
"It's unbelievable, it's really unbelievable." Alex quickly rushed to the door of the teacher's office. He didn't even bother to knock on the door, and pushed the door directly in. "Teacher, have you read it? The new year... .”
Before he finished speaking, Hinton slightly raised his right hand to signal him to stop.
After a long while, he slowly turned the chair around and said in a very low and deep voice, "This is too damn deep."
Google spent tens of millions of dollars to acquire Hinton's team, which is not a fan-chasing act. In their contract, Hinton's team plays a large role in helping them maintain their leading position in AI technology.
This time DreamNet has improved across generations, and the ten points at a higher position are far more valuable than AlexNet of my own team last year. Coupled with breaking the human standard, Hinton was suddenly under tremendous academic pressure that he had not experienced in decades.
For the first time in a year, he lamented that it turned out that Google’s money was so hot.
"Who are they? Is there any chance to contact them in advance?"
The Hinton team did not participate in the competition this year and did not participate in the hosting. Naturally, they did not have the email address and other information left when the team registered. Now that this shocking result has come out, there are too many people who want this information.
The organizer cannot give to everyone, so it is better to be fair and strict to the end.
Moreover, Xinton knew Li Feifei. She had just sent a message asking him to check the list, just to make himself curious, how could she provide assistance?
But Alex suddenly remembered something, "It seems that someone sent me an email a few months ago, and the signature was Meng, but I'm not sure if he has anything to do with the Dream team."
