"How about coming to study with me? I'll let you start directly from the last year of your undergraduate degree. After graduation, you will get a direct doctorate and I will give you a full scholarship."

Li Feifei sat in Han Ci's seat and started talking to Meng Fanqi.

"Hahaha, the full scholarship is not the key. I am mainly worried that I may not have much time in school in the next few years." Meng Fanqi smiled and said that Stanford's tuition is not cheap, about 50,000 to 60,000 US dollars a year, including living and accommodation expenses.

It costs US$70,000 to US$80,000 per year.

For ordinary students, this is a sky-high expense, and many American students need to borrow student loans. After working for several years, there is still a large amount of money that has not been paid off.

But hundreds of thousands of dollars is already a negligible figure for Meng Fanqi.

The income from the increase in Baidu's stock in the past few days alone has almost exceeded half a million US dollars.

Speaking of these stocks, Meng Fanqi plans to slowly sell these shares in batches in a few days and look for some potential companies to invest in the US stock market.

And he has already contacted Secretary Liu and transferred the second advance from Robin Li, totaling 18 million yuan, to a domestic account.

Because he suddenly remembered that at this point in time, it seems that there is still money in the country that can be picked up for free, and it is the last window period.

That is Mi Huyou, a company that was later valued at hundreds of billions and it seems that it has not yet seen the light of day.

"Let's go to Shanghai in the middle of this month." Meng Fanqi didn't remember the situation of several non-AI companies, and paid little attention to this aspect before his rebirth.

But Mi Huliu's experience is actually quite bizarre, and it's worth a special trip to Shanghai before he leaves.

"It doesn't matter if you're not in school, where is research not research?" Let's not mention that Li Feifei herself is closely connected with Google. If students can produce academic results of such quality, how can a tutor care so much?

"So good?" Meng Fanqi actually just got a degree as a token of his degree and didn't take it too seriously.

So many papers have to be published anyway, and the results have to be done, why not get a degree along the way?

As long as the tutor has less control, it will be fine. We don’t expect the tutor to provide much guidance and resources.

Putting aside his own unique advantages, he already had Hinton's guidance and Google's resources at this time, so his conditions in this regard were basically capped.

Li Feifei is obviously very aware of this, and what he focuses on is a relaxed, free and preferential approach, "You can come directly to the last year of undergraduate studies and earn your credits first. After graduation, you can directly come to me to study, and the articles will be published directly in cooperation with Google."

”

Not only is Li Feifei personally connected with Google, but he also directly served as Google’s vice president in 2017, mainly responsible for AI research and integration with cloud series products.

Even Google's two main founders, Sergey Brin and Larry Page, were both computer science doctoral students at Stanford.

Therefore, for both Stanford and Google, Meng Fanqi is working and researching at Google while studying for these degrees at Stanford. Our family does not need to make such a clear distinction.

This is also the main reason why Meng Fanqi named Stanford.

"Then it's agreed?" From Meng Fanqi's perspective, Li Feifei is a very good choice.

First of all, she is a Chinese immigrant and takes great care of Chinese people. For example, Jia Yangqing, who later became Ali's VP, was not Li Feifei's student when he was at Stanford, but he was urged to study by him many times.

Although his current status is not that high, he later became an academician of the three academies of the United States and was considered a first-class mentor in the Stanford AI circle.

Secondly, she has a very good relationship with top magazines such as Nature and Proceedings of the National Academy of Sciences, as well as organizations such as the three top visual conferences. Her IMAGE events are often held at top conferences.

By cooperating with her, it will be much easier to publish your results. With the quality of your results, you can basically say that you have booked awards for several conferences.

Finally, her influence at Google will make it easier to promote her results to a certain extent, and even the specific sharing and other internal matters at Google will go more smoothly.

Meng Fanqi had already thought about enrolling at Stanford directly at the venue, so he had already prepared a paper resume, research plan, and even a letter of recommendation written by Hinton.

It can be said that he was very prepared. Hinton could only sigh with regret as to why his school was so far away in Canada. As expected, those who were near the water had to be first-come-first-served.

At this time, Han Ci on the stage started his presentation after a lot of fuss.

"So far, the development of AI-related disciplines has completely changed people's past understanding of AI. Meng's residual ideas have achieved breathtaking achievements in many image tasks.

For example, identifying pictures more accurately than humans, or directly generating non-existent images out of thin air. These remarkable achievements are mainly accomplished through solving methods.

For example, for any image problem, we are interested in the mapping function from the image to the specific meaning, such as the category of the image to its content.

The current common training method is to give an efficient approximation of the objective function based on a limited data. Or it is to use limited samples without labels to approximate the unknown probability behind the selection."

"The basic components of neural networks are: linear transformation and one-dimensional nonlinear transformation. Deep neural networks are generally the repeated combination of the above structures.

For the already constructed network, design an optimization problem, fit the data based on empirical errors, sometimes add some regularization terms, and solve the optimization problem."

Han Ci's slide began to show some dense formulas, "From this, we can decompose the error into three parts. The approximation error is completely determined by the selection of the hypothesis space, and the estimation error is caused by the size and quality of the data set.

Additional errors, as well as optimization errors, are additional errors caused by optimization, or training."

Although I had a general understanding of Han Ci's thoughts in this regard, in such a formal occasion, I saw that she was a little nervous but still very confident in presenting these professional and profound contents.

Meng Fanqi still admires him very much.

Many of the conclusions here are well-known to the world later, but most of them are empirical conclusions summarized through a large number of experimental observations.

Although Meng Fanqi will break new ground in many AI application fields in this life, the inherent mathematical proof will ultimately be a gap that he cannot cross in his lifetime.

People will inevitably look at those who can do these things differently.

At this moment, in the venue, the one who was optimistic about Han Ci was not just Meng Fanqi, who was ignorant and thought he was powerful.

The academic group headed by Hinton, such as several old professors in the Oxford team, had eyes full of approval.

This kind of treatment was something that even Meng Fanqi didn’t get just now.
Chapter completed!