"This is so fucking impossible..." Yu Kai still couldn't understand this terrifying speed increase, "You must be using your newly proposed DreamNet here, right?"
"Yes, the residual network I proposed is very helpful in improving performance, and this help is universal. It is not only for classification tasks, but also for detection, segmentation and other types of tasks. Powerful performance improvements."
"But if you want to achieve this running speed, you actually can't use the 50 or 100 layers mentioned in your paper, right?"
"Of course, for fast detection algorithms, there is no need to use too deep a network structure. An 18-layer or 34-layer version is enough to meet most needs."
"No, this is impossible." Yu Kai took out a copy of the DreamNet paper brought back by Robin Li and carefully checked the parameters of DreamNet's 18th and 34th layers.
"For a network with this amount of parameters, you can only achieve 3-5 pictures per second at most." Yu Kai calculated over and over again, but still couldn't match the result at all.
Meng Fanqi originally wanted to explain that the huge speed increase came from innovative breakthroughs on the detection side, not the backbone network.
The YOLO method does not make a sliding window or propose optional areas, but directly performs regression on the entire image.
The generalization performance of this approach is very good, and the performance of pictures in different types and scenes does not fluctuate much, but it is slightly lacking in more detailed things, such as the detection and absolute position of smaller objects.
But as soon as I wanted to speak, I felt it was not safe.
The two technical leaders in front of me are both masters among masters. If they talk too much, they will make mistakes. If they wake up others, it will be a big problem.
"The specific results have been shown to several people. The details and principles of the algorithm are definitely not convenient to discuss with them at this stage." Meng Fanqi smiled and responded, "If you calculate it is impossible, it means that your premise is wrong."
"For me personally, there is actually a lot of room for improvement in this result, but my current main interest is not in this direction."
Listen, is this human language?
Wang Haifeng was about to say something, but he was speechless and choked on the spot.
He has developed a lot of detection algorithms in the past two years, and he is still far away from this result. What is the result of what the person in front of him said?
Not that interested in this direction right now? So how did you improve the detection accuracy while also speeding it up by more than a hundred times?
Just write a little casually, right? Are you angry?
"Will you be responsible for continuing to optimize this series of algorithms in the future?" Robin Li is very concerned about this matter. If Meng Fanqi agrees to continue to optimize and upgrade this series, it will actually have an effect similar to recruitment.
"Then it depends on how we sign the contract." Meng Fanqi started Tai Chi. He didn't see the contract, so it was difficult to say anything like this.
Li Yanhong leaned back on the chair, rested his chin with his left hand, and began to think deeply.
Meng Fanqi does not doubt Li Yanhong’s investment in AI. In the ten years from 2013 to 2023, Li Yanhong has invested more than 100 billion in AI, averaging tens of billions every year.
Even if one percent is allocated from the annual budget, it will be enough for oneself to have enough to eat.
"We also have some picture data here, can we move it over for some reasoning?" Wang Haifeng asked.
"No problem, everything is fine." Meng Fanqi suddenly became alert after hearing this. Generally speaking, asking an algorithm to reason directly on its own data sounds normal, but in fact it is not very reasonable.
The data is different, and the types in the pictures are likely to be completely different, so naturally they cannot be detected.
Corresponding training data is needed to fine-tune the model to be more reasonable.
Combined with the questioning attitudes of the previous two technicians, Meng Fanqi began to wonder if he was suddenly called here today because someone simply didn't believe his results.
Although I feel a little unhappy, it’s understandable.
Meng Fanqi took out an external camera directly from his bag, "Or you can just connect a camera directly. We won't spend that effort moving the data."
There is a certain risk in plugging a USB flash drive directly into a computer, which is why many major manufacturers later did not allow employees or other personnel to connect any external devices to their hosts.
The repeated questions from Baidu's two technical staff made Meng Fanqi's response cautious.
The previous communication with Robin Li was so smooth that my previous mentality was a bit childish. I still need to be more cautious when dealing with such a major transaction.
"Have you done relevant tests in advance and added interfaces?" Yu Kai felt completely unsure.
Connecting an external camera is the most direct and crude method. Everyone can see the effect of the detection algorithm on the content captured by the camera in real time.
It is almost impossible to fake this thing.
Just now, my eyes indicated that Wang Haifeng proposed to use Baidu's own data for testing. In fact, the subtext was that Meng Fanqi may have used this part of the test data to fine-tune his own model in advance.
To put it bluntly, it is cheating, allowing the model to learn the data that will be used for testing first. After reading the reference answers and then answering the questions, your score will naturally improve by leaps and bounds.
And connecting an external camera to take real-life measurements is equivalent to a third-party examiner giving questions on the spot, and there is no chance of cheating at all.
Since he had already done testing and adaptation before, it didn't take long for Meng Fanqi to connect the camera and start running his own algorithm.
The three senior executives pointed the camera at Baidu. The image on the computer screen was quickly framed by an algorithm to select the positions and categories of people, tables, chairs, computers and other elements.
Meng Fanqi deliberately shook the camera, and all the selection boxes were almost close to the target object, following smoothly. There is absolutely no problem with the current detection algorithm, which is that the detection frame cannot catch up with people.
Meng Fanqi held up the camera and took pictures everywhere, and found no problems with the recognition of common objects, such as books and water cups.
At this moment, even if they can't think clearly anymore, the two technical leaders are still materialistic and believe in science.
Yu Kai took a deep breath and said, "This is a very terrifying breakthrough..."
"I wonder what the value of this 'terrible breakthrough' is from the management perspective of an Internet giant?"
To be honest, Meng Fanqi really didn't know about this matter. He knows the details of these technologies and the power of breakthroughs.
But if the timeline is slightly advanced, it is difficult for Meng Fanqi, who lacks high-level management experience, to estimate how much value this thing can bring to an Internet giant like Baidu.
