Find the original at bit.ly/3iBfjkV.

"Therefore, I believe that the lack of better results in deeper networks is an optimization problem, rather than a model design problem or a model capability problem. The model itself has greater potential, but the optimization method needs to be changed."

"And this is the significance of deep residual."

"For any mapping H(x) that we need to learn, I expect the network to learn the F(x) in F(x)+x=H(x), rather than directly learning H(x) itself."

"This operation can be simply implemented by adding an addition, and the F(x) with a difference of x from H(x) is called a residual mapping with respect to identity."

"If this identity is ideal, then we can easily set the weights to a very small value. The form of the residual effectively solves the gradient problem that has always been troublesome. As you can see, it allows networks with hundreds of layers to continuously achieve better performance."

"The achievements in the recognition classification of this competition are just the basic embodiment of the residual idea. In fact, it is a better feature extractor that can better extract image features for various image tasks."

"Not only in the detection track of this competition, but also at Baidu's technical conference a few days ago, we have seen obvious performance improvements."

At this point, Meng Fanqi paused because the audience below had already started discussing.

Although Baidu announced that the real-time detection method was mainly contributed by Meng Fanqi, a special researcher, they did not disclose any specific algorithm details. They only promised to announce them 6-12 months later.

Although most people have probably guessed it, hearing Meng Fanqi say it himself at least confirmed one thing: Baidu is now far ahead in real-time detection algorithms and is using DreamNet.

"Of course, there are still many visual tasks. In addition to the ones I have already done, such as generation, detection, segmentation, and recognition, there are also pose estimation, depth estimation, super-resolution, and so on. There are many different types."

"My DreamNet paper has been published and the code has been open-sourced. There are still many directions that we need to explore together."

Meng Fanqi has already taken the largest share of the cake in several major tracks, so naturally he needs to leave some opportunities for others.

In a more specific direction, a few highly recognized representative works are enough.

Meng Fanqi has so many technical skills that there is no need for him to do everything in every subfield.

In other words, it is just a matter of slightly modifying things to make the machines run, changing the data, and adjusting some individual structures and parameters.

By open-sourcing the code, more and more work can be based on his technology and algorithms, which is a more cost-effective approach.

Meng Fanqi finished speaking, actually only taking about ten minutes in total.

According to the original plan, he could have spoken for about 25-30 minutes.

Unfortunately, he became arrogant and his attitude changed after receiving the advance payment from Li Yanhong. Meng Fanqi no longer has the psychological need for recognition and approval from the academic community.

Looking back, Meng Fanqi was a little worried before the advance payment from Li Yanhong arrived, feeling a bit uneasy and hoping for recognition.

He doubted how much resources he could really leverage.

Now, all of that is in the past. Without the need for recognition, Meng Fanqi's presentation has become more concise.

Ten minutes is actually not short in this kind of occasion, especially since there are two parts to this presentation. After Meng Fanqi finishes, there will be Han Ci's theoretical explanation.

Therefore, everyone in the audience did not find it unusual, except for Han Ci, who was stunned. What's going on? Wasn't there supposed to be ten more minutes? Why is it my turn already?

"While doing this research, I received a lot of help from a mathematics professor at our school, Fu Deqing. He is a co-author of the paper, but not a scholar in our field, so he has no intention of participating in this conference."

"For the reason why the residual idea is effective and what its significance is, we have invited Fu Professor's junior sister, Han Ci, to bring us her perspective from the perspective of dynamical systems."

After Meng Fanqi finished speaking, he prepared to walk down to the stage, took two steps, and then turned back to the microphone to add.

"By the way, since I have signed with Google and considering my current academic situation, I urgently need a university near Silicon Valley to accommodate me."

"I hope Stanford can be more accommodating."

After speaking, the whole audience burst into laughter.

The teams participating in this competition are from Microsoft, UC Berkeley, St. Petersburg, IBM, Donghua University, National University of Singapore, Oxford, and University of Toronto.

The level of the attendees is also high, mostly holders of master's and doctoral degrees from top-tier universities and tech giants.

These people's results have been completely surpassed. They have gathered here to listen to Meng Fanqi introduce his algorithm, which has made many people who don't know the inside story completely ignore Meng Fanqi's current undergraduate student status.

Just based on the quality and level of the papers he has already published, they are enough to meet the standards of a doctoral graduation.Network structure, generation, segmentation, optimizer, and normalization methods, anyone with discerning eyes can see that these ideas will become the basic paradigms of the new AI era.

There are even two or three groundbreaking and pioneering works in new directions.

No one ever thought that he still had such problems to solve.

After Meng Fanqi finished speaking, one of the data preparation personnel for the competition and one of the mentors of the Stanford AI Lab, Li Feifei, immediately started on-site recruitment.

A few years ago, when she was collecting IMAGENET data, due to the huge gap between the capabilities of the algorithm and human level, Li Feifei always hoped that one day AI algorithms could surpass humans on the large-scale data she collected.

She originally thought it would take ten to twenty years, but it turned out to take less than five to six years.

Especially in the past two years, the accuracy has been explosively improved by 20%, directly fulfilling her wish.

Now that her wish has been fulfilled, and the person who accomplished this happens to be seeking the opportunity to study at Stanford, Li Feifei naturally wouldn't let it pass.

Moreover, in 2014, Stanford was just about to start offering courses in deep learning, and Meng Fanqi's joining would greatly benefit this.

Li Feifei thought so, completely unaware that her idea seemed a bit strange, recruiting an undergraduate student, but thinking of letting him assist the university in providing course quality.

It's really unclear whether she's recruiting an undergraduate student or a lecturer, this situation seems a bit absurd.

There were quite a few professors and scholars from prestigious universities such as Oxford, Cambridge, MIT, and Harvard present. They originally heard that he wanted to study in the United States and had the intention to recruit him.

But when they heard that Meng Fanqi had signed a contract with Google and specifically mentioned Stanford, a few old professors who were somewhat burdened felt a bit awkward.

Seeing Li Feifei and Meng Fanqi chatting and laughing, they didn't feel comfortable to interrupt.