"Go to the Baidu headquarters in advance?" Meng Fanqi was stunned for a moment, "It's mainly because I have something to do. I've been free for the past two days, so it's convenient to go there. In fact, everything is almost ready."
"Are you almost ready? That's great. Please bring your things with you when you come. Our main technical person here wants to see you."
"If we have these contents, we can test them in the actual framework and server first, so that we can have a better idea of the specific actual usage performance, and we can continue our subsequent work."
There is no problem in Robin Li's statement, but in fact, there are some inconsistent voices within Baidu's AI Research Institute.
"No problem, I have some things to do here now. After the time is determined, I will contact Mr. Li to confirm the specific time." Meng Fanqi didn't think too much. In his opinion, it was probably because of App-related development matters.
Baidu needs to confirm the speed of their algorithm in their environment and equipment so that they can make many specific plans.
How much detection speed can be achieved, how much resources are occupied, and if it is run on an iOS or Android device, how will the speed change?
The workload of these matters is still quite large. Meng Fanqi believed that Robin Li put too much pressure on his team, so the technical person in charge there would think of contacting him as soon as possible.
He had no idea that Baidu's technical director was skeptical about the authenticity of his results.
Li Yanhong is the founder and president of Baidu AI Related Research Institute. Although he was born with a master's degree in computer science, he has returned to China to start a business for more than ten years, so it is impossible for him to always pay high attention to technical details.
Li Yanhong actually learned about AI and deep learning-related content from Yu Kai, a returnee executive who worked at Baidu in the past two years.
Yu Kai was already very concerned about neural networks when he was an undergraduate at Nanjing University, which was less than 2000. In the following ten years, he continued to work in this field.
After joining Baidu in early 2012, although Li Yanhong nominally served as the dean of Baidu Deep Learning Research Institute, the actual person at the helm and the main person in charge of technology was Yu Kai, the executive vice president.
Yu Kai has deep experience in the industry. Since 2012, he has been mainly responsible for promoting the application of deep learning technology at Baidu.
Many of Baidu's technological breakthroughs in speech recognition, image recognition and advertising recommendation in the past two years were led by Yu Kai.
In the original timeline, Andrew Ng, who shocked the world in 2014, joined Baidu, largely because of the persuasion of Yu Kai, Ng's friend for many years.
Last year, when competing with Google for talent, Robin Li sent this capable person.
Under the influence of his influence, quite a few top technical talents joined Baidu during this period.
Based on his research experience, the experimental direction given by Meng Fanqi happens to be the professional part in which he is best at.
Yu Kai won the championship in the international visual recognition competition in 2009, including IMAGENET, which Meng Fanqi won this time. Yu Kai also led the team to win the first competition in 2010.
It’s just that neural networks had not yet received attention at that time.
In Yu Kai's view, the results brought back by Robin Li some time ago were quite unreal.
"Stanford, MIT, Harvard, any top university I haven't studied at? What is the level of the world's top detection algorithms now? Is there anyone who knows better than me?"
The logic of Yu Kai's question was not unreasonable, but the tone was indeed a bit harsh.
On the one hand, it is because this is my professional field, so I am inevitably a little unwilling to admit defeat. On the other hand, Li Yanhong's unconditional trust in this exaggerated result also made Yu Kai feel a little unhappy.
If this result is truly effective, once it is put into use directly, it means that a lot of related work in the past one or two years will be wasted and become useless.
Instinctive rejection is also human nature.
"Boss, are you deceived? He single-handedly developed a technology that is hundreds of times more accurate and faster than the latest achievements of our research institute?"
If this technology is real, Yu Kai kind of hopes it's fake. But if it was true, Yu Kai actually wanted it to be true.
What an ambivalence.
"Lao Yu, your perspective and ideas seem a bit narrow. Although we did not exchange specific technical details that day. But judging from his two huge breakthroughs this year, there is no doubt about the authenticity of this matter."
At this time, Li Yanhong also felt that his way of handling this matter might be somewhat inappropriate. Directly bypassing the dedicated technical staff, which may make them a little emotional
But it was indeed relatively sudden at the time. At first it was just to recruit talents, and there was no plan for technical cooperation.
"I just read his paper carefully, and I feel that this matter makes no sense." Of course, Yu Kai does not deny the achievements of Meng Fanqi's DreamNet and generative adversarial networks.
These two results have now been concluded, and the general response and evaluation in the industry are very high.
Yu Kai speculated that the results that Li Yanhong got back used the neural network of DreamNet technology as the backbone, which was why there was such a big breakthrough in detection performance and accuracy.
He also read carefully the DreamNet papers that Li Yanhong brought back.
"DreamNet makes the neural network very deep. It adopts a very simple and elegant structure and solves the difficult problem of deep network optimization."
Yu Kai was very amazed and appreciative of this result.
"But at the same time, your network will also be deepened several times or even more than ten times. In this case, the parameters and computational load of the network will of course increase exponentially.
Of course, some processing and structural adjustments can be made to balance this problem, but in the final analysis, it is good that the detection speed does not slow down. How can it be accelerated by hundreds of times? "
This is actually a misunderstanding among ordinary people. They feel that if a person has just made some very important breakthroughs, it will be difficult for him to make huge breakthroughs in other areas at the same time.
In Yu Kai's view, Meng Fanqi directly opened up two major fields in half a year, unsupervised generation, and the true "deep" learning era.
He will not have the time and energy to conduct in-depth research on detection algorithms at the same time.
In addition, what Li Yanhong brought back this time was a version of DreamNet's manuscript, which also misled Yu Kai's thinking.
He mistakenly believed that Meng Fanqi's detection algorithm technology breakthrough mainly came from the application of DreamNet.
So he thought he quickly understood the reason for the rapid improvement in performance. However, after reading the DreamNet paper and slightly calculating the parameters of this structure, he could not understand where the speed improvement came from.
"Either this, I'll invite him over, and we'll have a technical exchange first. This way, other related aspects can be easily adapted."
Li Yanhong suddenly contacted Meng Fanqi. On the one hand, he wanted to dispel the doubts of his subordinates. On the other hand, he himself was said to be a bit uncertain.
