"DreamNet? What kind of network structure is this?" Li Feifei carefully searched through the descriptions submitted by each team.

Almost all of the submissions from various teams described their structures as "based on AlexNet," "modeled after AlexNet," or "combining AlexNet with traditional support vector machines/.../." The methods were similar, and the results were naturally not much different.

The remaining descriptions focused on how to integrate multiple models together.

When Li Feifei came across the team "Dream," she noticed that the personnel column only listed two names, "Fanqi Meng" and "Huang Tang," without any specific organizational information.

The introduction to the team's algorithm was also brief.

"We propose a new DreamNet that allows deep networks with hundreds or even thousands of layers to be trained. Due to a lack of sufficient resources, we only obtained a single set of weights for submission."

Meng Fanqi did this intentionally. Of course, he could have briefly explained the residual or other major tricks in the team's submission description. But why did he choose to do it this way? He had to first arouse the curiosity of the industry professionals in order to gain the maximum attention.

Top-5 classification error rate: 4.9%.

This simple number struck Li Feifei's heart directly.

The algorithm's error rate was lower than the human error rate of 5.1%. Since she started promoting this competition in 2009, it had always been her goal and a major reason for organizing IMAGENET. She even made up her mind that if an algorithm in a certain year successfully surpassed this benchmark, she would no longer continue organizing IMAGENET.

But she never expected that this day would come so soon.

Just 400 days ago, the best traditional methods could only achieve an error rate of around 25%.

Last year, AlexNet achieved an error rate of around 16%, which she thought was a significant leap forward. Such progress was considered a once-in-decades occurrence.

We are bit.ly/3iBfjkV, find us on google.

She never imagined that this year's results would bring an even more astonishing surprise.

DreamNet, surpassing the human benchmark.

Li Feifei felt that the name was very fitting. The result was truly like a dream.

Detection result mAP: 0.34673, dealing a heavy blow to the second-place result of 0.22581.

My goodness, who are these people? Li Feifei racked her brain, but she couldn't recall hearing these names in any AI laboratory.

"Why not provide a general description of the algorithm?" Li Feifei pounded the table twice with her hand. As a data collector, no one understood and appreciated the value of algorithm performance more than her.

Since she saw this result, her curiosity had been uncontrollable, like a wild horse. She really wanted to know, even if it was just a few sentences of vague description or a couple of specific terms that were hard to understand.

Unfortunately, Meng Fanqi was well-versed in the art of media promotion. In his submission description, he first stated, "We propose a new DreamNet," indicating that this structure was completely different and had undergone significant changes compared to AlexNet, arousing the readers' curiosity.

Then he dropped a bombshell, "We enable deep networks with hundreds or even thousands of layers to be trained," comparing it to last year's champion with only 8 layers. This number was too shocking and made the technical practitioners itch with excitement.

Finally, he pretended to be modest, but actually implied that they only used limited resources to achieve this result, without going through the process of integrating multiple models like others.

The performance difference was so significant that it was unnecessary.

"Get the results page online as soon as possible." Li Feifei replied to Deng Jia's message, thinking to herself that it was no wonder Deng Jia, who was usually calm and steady, was so excited. "I can't bear this burden alone. I want to make all of you as anxious as ants on a hot pan."

The top minds were always filled with curiosity. This curiosity about the unknown and the pursuit of something better drove them to achieve one breakthrough after another that could change human life.

For results that were somewhat similar, one might still be able to tolerate them. Because they might think that if they thought carefully, tested diligently, and added some small techniques, they might be able to achieve the same results.

But when faced with an improvement that was even more astonishing than last year's AlexNet, even the most arrogant person had to admit that they couldn't achieve this in a short period of time.

The greater the gap, the less room there was for doubt and jealousy, only pure curiosity and admiration.

Meng Fanqi didn't plan to change the names of most of the techniques, but he specifically named the residual network trained with other methods such as batch normalization and second-order optimizers as DreamNet.

One reason was that he had been reborn for ten years, and going back to 2013 and submitting a residual network felt like a dream. He occasionally worried that this dream would wake up at any moment.

The other reason was that his surname was Meng, which sounded like "dream" in Chinese. If he named it MengNet, it would be too thick-skinned and unnecessary.

After the competition results came out, Tang Huang and Liu Xu also felt like they were in a dream.

Even though Meng Fanqi's code updates and small tricks consistently improved the model's performance and training speed, they still wondered if they could really win the championship.

But deep down, they only had doubts and never truly believed in this possibility.

Not to mention winning with such a crushing attitude.

"Microsoft, UC Berkeley, St. Petersburg, IBM, Tsinghua University, National University of Singapore, Oxford, University of Toronto." As Li Feifei listed these names one by one, she realized that many world-class universities and institutions were left far behind.Even Li Kun, one of the three AI giants, personally led his team to participate in this competition, but he still ended up defeated.

Meanwhile, as Li Fei Fei urged Deng Jia to announce the results on the webpage, three people in Google's Mountain View office in Silicon Valley near Stanford University were the first to notice the results.

They were last year's champions, the Hinton team, which Google had acquired for a hefty sum.

Alex, tall and thin, with short hair and a square face. Dressed in the classic programmer's outfit, his actions were crisp and efficient, giving off an aura of wisdom and competence.

Hinton had a three-dimensional face, a slightly hooked nose, and bore a slight resemblance to Snape. When his face was expressionless, he appeared very imposing, instilling fear in others, but when he smiled, he was very infectious.

However, at this moment, he couldn't muster a smile.

"Unbelievable, truly unbelievable." Alex rushed to the door of his teacher's office, not even bothering to knock, and barged in, "Professor, have you seen it? The new year's...."

Before he could finish his sentence, Hinton slightly raised his right hand to signal him to stop.

After a long while, he slowly turned his chair around, his voice very low and deep, "This is really too profound."

Google had spent tens of millions to acquire the Hinton team, not as a fan chasing stars. In their contract, the Hinton team's assistance in maintaining Google's leading position in AI technology accounted for a large part.

This time, DreamNet's epoch-making improvement, the higher position of ten points, was far more valuable than their own AlexNet last year. Coupled with breaking human standards, it suddenly gave Hinton an academic pressure he hadn't felt in decades.

For the first time in a year, he felt that Google's money was burning a hole in his pocket.

"Who are they exactly, is there a chance to contact them in advance?"

The Hinton team did not participate in the competition or co-host this year, so naturally, they did not have the email and other information left when the team registered. Now that these shocking results have come out, there are too many people who want this information.

The organizers couldn't possibly give it to everyone, so they might as well be fair and strict to the end.

Moreover, Hinton knew Li Fei Fei, she had just sent a message asking him to check the leaderboard, obviously trying to pique his curiosity, how could she possibly help?

However, Alex suddenly remembered something, "A few months ago, someone sent me an email, signed as Meng, but I'm not sure if he has any connection with the Dream team."