“DreamNet？ What is this network structure? Li Feifei carefully rummaged through the descriptions submitted by each team.

You can see that almost all of the team's submissions describe, "Our structure is based on AlexNet", "We model AlexNet", "We combine AlexNet with traditional support vector machines/.. /.. combined".

The approach is similar, and the results are naturally very different.

All that's left is a description of how to put several models together.

Flipping to Team Dream, there are only two names in the personnel composition column: "Fanqi Meng, Huang Tang", and there is no specific organization information.

Watch the introduction of the team's algorithm, and there are even fewer words.

"We have come up with a new DreamNet that enables networks to be trained at depths of 100 or even 1,000 layers. Due to the lack of sufficient equipment, we were only given a single weight for the commit. ”

Meng Fanqi did it on purpose, and of course he could have briefly described the residuals or other major tricks in the team's submission description. But why do you do this, you must first arouse the curiosity of these people in the industry before you can get the most attention.

Classification top-5 error rate, 4.9%.

Such a simple number directly hit Li Feifei's heart.

The algorithm has an error rate of less than 5.1% of humans, which has been her goal since she promoted this contest in '09, and it is one of the reasons why she organized this contest.

She even planned that if a certain algorithm succeeded in breaking through this level, she would not continue to organize IMAGENET.

But she never imagined that this day would come so quickly.

Just 400 days ago, the best traditional approach had an error rate of about 25 percent.

Last year's error rate of around AlexNet 16 was a leap forward that she thought was a once-in-a-decade progress.

I didn't expect this year's results to give her an even more exaggerated surprise.

DreamNet, the DreamNet that breaks through the human level.

Li Feifei felt that the name was really appropriate, and this result was really dreamlike.

The test result mAP: 0.34673, which is a dimensionality reduction blow to the second place 0.22581.

My God, who are these people sacred? Li Feifei pondered and meditated, and couldn't think of any AI lab where he had heard these names.

"Why don't you give a general description of the algorithm!" Li Feifei slammed the table twice with his hand, and no one better understood and understood the value of the algorithm's performance than her, the data collector.

From the moment she saw the result, her curiosity was like a wild horse on the loose, and she could no longer control it. I really want to know, I want to know. Even if it's just a plausible description or two, even if it's just a specific term or two that doesn't really matter.

It's a pity that Meng Fanqi is well versed in media propaganda, and between the lines, he first said "We propose a new DreamNet" to indicate that this structure is completely different, and it has changed a lot compared to AlexNet, so as to strengthen the reader's curiosity.

Then a bombshell was thrown, "We make the network up to 100 or even 1,000 layers deep can be trained", compared with last year's 8-layer champion, this number is too shocking, and it makes technology practitioners itch.

Finally, I pretend to be modest, but in fact, we only use relatively limited resources and get a result, so we won't go with you to roll up the multi-model integration set.

There is so much difference in performance, it's really unnecessary.

"Get the results page online as soon as possible." Li Feifei replied to Deng Jia, thinking in his heart that it was no wonder that Deng Jia, who was calm and down-to-earth, was very reckless just now, "I can't let me suffer this crime alone, I want to make you all anxious like ants on a hot pot." ”

The top brains are very curious, and this curiosity about the unknown and the pursuit of the better drive them to make one achievement after another that can change human lives.

It may be easier to be patient with slightly more similar results. Because I feel that if I think about it carefully, test it hard, and add a little skill, I may be able to reap the same result.

And an even more exaggerated improvement than last year's AlexNet is in front of us, and even the most conceited people have to admit that they can't do it at all in a short period of time.

The greater gap makes people less suspicious and jealous, only more pure curiosity and admiration.

Meng Fanqi does not plan to change the name of most of the techniques, but only calls the residual network results trained by some other methods such as batch normalization and second-order optimizer as DreamNet.

First, because he was reborn in a dream for ten years, his return to 2013 to submit the residual network was like a dream, and he actually occasionally worried about whether this dream would wake up at any time.

The second is because his surname is Meng, which is homophonic with dreams. If you want to really name MengNet submission, you will be too thick-skinned and there is no need for this.

After the results of the event came out, Tang Juan and Liu Xu also felt that life was like a dream.

Even if Meng Fanqi's code and tips can always improve the performance of the model or the training speed every time they update, even if they have wondered countless times, they can't really win the championship, right?

But deep down, they were all just skeptical, never really believing in it.

Not to mention winning the championship with this crushing attitude.

"Microsoft, UC Berkeley, St. Petersburg, IBM, Tokyo, National Singapore, Oxford, Toronto." Count down how many of the world's leading universities and institutions have been left far behind.

Even Likun, one of the AI triumvirate, personally led the team to participate in the competition this time, but he could only sink into the sand.

At the same time, at the same time that Li Feifei asked Deng Jia to announce the results quickly, three people in Google in Mountain View, Silicon Valley, near Stanford University, immediately paid attention to the results.

They are last year's champions, the Hinton team, which was bought by Google with a lot of money.

Alix, tall and thin, with short hair and a square face. Dressed in a classic programmer suit, his words and deeds are crisp and neat, and he has a very intelligent and capable feeling.

Hinton has three-dimensional features, a slight hooked nose, and looks like Snape. When the face is expressionless, it looks very majestic and intimidating, but when it laughs, it is very infectious.

But at this moment, he really couldn't laugh.

"Unbelievable, really unbelievable." Alix hurried to the door of the teacher's office, and didn't bother to knock on the door, so he pushed the door and entered, "Teacher, have you seen it?" New Year's...."

Before he could finish speaking, Hinton raised his right hand slightly and motioned for him to stop.

After a while, he slowly turned the chair around, and said in a very low and deep voice, "This is too deep." ”

Google's spending tens of millions of dollars to acquire the Hinton team is not an act of fans chasing stars. A large part of their contract was for the Hinton team to help them stay ahead of the curve in AI technology.

This time, DreamNet has improved across the era, and the gold content of ten points in a higher position is far higher than that of AlexNet in his own team last year. In addition, it broke the human standard, and suddenly gave Hinton a huge academic pressure that had not been seen for decades.

For the first time in a year, he had this kind of sigh, it turned out that Google's money would be so hot to hold.

"Who the hell are they, and do you have a chance to get in touch with them in advance?"

The Hinton team did not participate in the competition or participate in the hosting this year, so naturally there was no email address and other information left when the team registered.

It is impossible for the organizer to give everything, so it is better to be fair and strict to the end.

And Hinton knows Li Feifei, she just sent a message to let her check the list, just to make herself curious, how could she possibly help?

Alix suddenly remembered something, "It seems that someone sent me an email a few months ago, signed by Meng, but I'm not sure if he has anything to do with the Dream team. ”