Back at the hotel, Wang Yinan opened the laptop he was carrying, activated FR ID, and began to organize his thoughts.

Artificial intelligence is the current hot spot, and neural networks are the hot spot among the hot spots, but except for a very few fields, such as license plate recognition, or a very few cases, such as Mi Ge's Go dog, there are few specific application cases in practice.

The reason for this is very simple, that is, the dependence of neural networks on hardware, it is not impossible to run neural networks on ordinary computers or mobile phones based on AR chips, but the efficiency of operation will definitely drive you crazy.

And with the huge amount of data in the neural network, it is almost impossible to install it all in the mobile phone.

So, the current neural network application is extremely dependent on hardware, which limits its application range, anyway, you keep in mind that it is almost impossible to run a practical neural network on an ordinary mobile phone.

But with Wang Yinan's new discoveries and new technologies, it is completely different, there is no shortcut to the sorting and training of samples, and it is good to spend a long time on special hardware honestly, which is not the focus of the problem.

But once the training is complete and the neural network is stabilized, Wang Yinan can use a series of mathematical tools and formal skills to block the huge neural network data according to the type of neural network.

To put it simply, it can be thought that the multi-layer iterative operation mode of neural networks can be mapped to the operation mode of ordinary von Neumann machines.

Specifically, for example, for the neural network chunks, map them to a generic von Neumann machine instruction, e.g., java or llv bytecode.

You can think that Wang Yinan has found a compilation method that can compile the data of the neural network into a general bytecode, and then use the ready-made JI (just-in-time compilation) technology, or LLV compilation technology, to obtain an executable P or mobile phone program.

Finally, by combining all the chunks, you can get a compiled version of any trained neural network, either on p or on a mobile phone.

That is to say, the biggest obstacle restricting the development of artificial intelligence, the limitation of hardware is now completely non-existent, as long as the neural network that meets the requirements is trained with enough samples on the dedicated hardware, and then Wang Yinan's technology is used, the trained neural network can be compiled to an ordinary P or even a mobile phone, so as to get a mobile phone app that can complete the same task.

Let me give you another specific example, and you will understand that at present, Ge Kunpeng 2.0 has learned how to fly with the highest efficiency through training, but it cannot directly teach humans how to fly, and it cannot take Ge Kunpeng into the sky and directly let it command how to fly.

Therefore, the method adopted by Wang Yinan and Chief Master Yang now is to write the flight control software by themselves, and then let Ge Kunpeng find the defects inside, and improve the flight control software written by humans after many iterations, so as to obtain good quality flight control software.

With Wang Yinan's new technology, it is now possible to directly compile the neural network of Ge Kunpeng into the instruction set of the flight control computer, such as the alpha system, and the compiled program is actually the best flight control software!

It doesn't matter if you can't understand or describe it, just execute it!

Even if you add a vector engine to the new model's record, it will only take two more days of training and recompilation, and you will immediately get the best flight control software.

"This is a big game", Wang Yinan finished typing the above words on the computer, rubbing his hands, not knowing what to say.

Seriously, if IA, FBI and the like knew what Wang Yinan was doing now, it was estimated that even if they sent all their agents in the whole of China, they would kidnap him.

Of course, it will take a lot of time and work to complete this idea, sorting out all the types of neural networks, the corresponding mathematical tools, and generating abstract instruction sequences for different data.

Neural network data that is only formalized using mathematical tools must be lengthy, and in order to achieve practical use, the code needs to be streamlined and optimized, and of course, the Gödel system can play its best in this regard.

It's best to be able to automate and program the whole process, but this is not in a hurry, human intelligence and machines to work, has always been the fastest and most efficient solution.

Wang Yinan dispelled the idea of finding the first place to eat crabs, although this will definitely be very cool, but the latest stealth fighter of the institute is not to mention Huaguo, how many pairs of eyes are staring at the entire blue star, if he accidentally makes an earth-shattering flight control, Wang Yinan can think of how big the movement is with his toes, obviously this is definitely against his original intention of keeping a low profile.

It's better to make a fortune in a muffled voice, Wang Yinan focused his attention on David's plan, or to be precise, a point in David's plan, you know, the biggest problem with David's plan now is the lack of a killer-level application, and the kind of functions and effects that far surpass others.

The basic chat function, no matter how difficult it is, is difficult to play, and only in some areas that are not very mature now, such as multiplayer video chat.

At present, the mainstream of video chat technology is H24, which is also the video codec technology supported by most mobile phone hardware, for one-to-one video chat, as long as the requirements for resolution are not too high, the effect is passable.

Of course, if you want to work well, you also need newer technologies such as H25, or Mi Ge's VP0, ah no, because of H25's patent issues, a bunch of industry leaders got together to open a new project, called the Open Media Alliance, they made an adult film No. 1 format, no, it's audio and video No. 1 format (AV), which is claimed to be more efficient than H25.

Whether it is H24, H25 or AV, the video is compressed according to the traditional method, or the continuous image stream, that is, the video is processed on the spectrum to reduce the parts that the human eye is not paying much attention to, and then through the analysis of the time domain, the duplicate parts between adjacent image frames are removed, etc.

Of course, in between, chunking technology is used to reduce computational complexity, or adaptive chunking technology is used to automatically make trade-offs between coding efficiency and coding quality.

Wang Yinan decided to take a different approach and approach the problem exactly like a neural network, just like Mi Ge's version of Doggyr, who discarded all prior knowledge, even the concept of frequency.

The only input source of the neural network is the sequence of images, and the quality criterion of the evaluation algorithm is the deviation between the decoded image sequence and the original image, which is a method for evaluating the variance of human vision.

Using this method to train the neural network, Wang Yinan takes the algorithm time as a basic parameter of training, that is, the smallest size, the fastest time, and the smallest visual variance of the human eye, and trains with all the samples that can be obtained.

After Wang Yinan returned to the imperial capital, he briefly met with the team of the David Project, looked at their prototype, and then asked the team to start developing version 2.0 for the problems of the prototype system, focusing on the extensible architecture and the stability of the service.

And he himself began to retreat again, the school is obviously not a good retreat place, let alone Zhou Hui's family, Wang Yinan locked himself in the base camp of Ronghe company, so it is okay, closer to the Gödel system.

Wang Yinan during the retreat was like a completely changed person, his six relatives did not recognize, his eyes were dull, and he was completely incompatible with the usual civilian-looking young man, and the whole was the image of a mad scientist, but the efficiency was really amazing.

Zhou Hui couldn't help but go to Ronghe to see him twice in the company, but each time he couldn't stay for a while and ran away in embarrassment, in her words, it was too crazy, Wang Yinan sat in front of the keyboard most of the time typing frantically, sometimes floating around in the office like a ghost with no focus on his eyes, and there were words in his mouth, which was really scared to death.

The Gödel system began to use countless video samples to start training, what yub, what love art, what is cool, anyway, Wang Yinan bought countless VIP accounts for the Gödel system, and frantically let it watch videos, as for whether it will go crazy, it is not the category that Wang Yinan considers.

But the number of these video samples doesn't seem to be enough, the key is that basically all the materials have undergone various post-processing, or beautification, or editing, anyway, for neural network learning, it's not original, and the noise is a bit too much.

At this time, Wang Yinan can't help but thank Principal Wang for the live broadcast trend, in addition to finding a crazy red sea for venture capital, and providing a channel for some two-dimensional dicks to get rich for men and women, countless live broadcast software has also provided countless real video samples to the Gödel system.

So at Ronghe headquarters, countless live video streams were brought in, and then converted into a dedicated neural network weight for video compression in the Gödel system.

It's a neural network that has abandoned all prior knowledge, no Fourier analysis, no wavelet analysis, no motion prediction, no chunking, and no discrete cosine transformations.

The goal of the neural network is also very clear, to compress the raw video data, or discrete image frames in the time domain, with the smallest possible size, as fast as possible, and as little variance as possible to the human eye, as for what the principle is, who cares?

Wang Yinan's task is very simple and heavy, he has to find a series of mathematical tools, formal techniques, and program these so that the Gödel system can compile the trained neural network into llv bytecode.

After many iterations and optimization, a condensed version of the LTV bytecode was finally obtained.

These bytecodes can be located in a specific PU architecture, such as AR, to get a compression and decompression program that can actually work.

I don't know if the Gödel system will learn Northeast dialect after watching too much live broadcast, I don't think it will be, but Wang Yinan resolutely did not let the Gödel system use that short video of shaking to train, scientists also have a habit of cleanliness, "Then l things, pollute me, don't pollute the Gödel system", Wang Yinan thought.

Time flies, during this period, Wang Yinan also went to two classes and watched two movies with Zhou Hui, after all, a mad scientist is still a normal person.

The video compression neural network of Gödel's system has gradually stabilized, and the work of compiling to llv instructions has not yet been completed, Wang Yinan can't wait to see the effect of directly using the neural network for video compression, after all, if there is no H25 strong, it is better to wash and sleep.

He chose a Blu-ray 4K resolution film source, which is the kind of short video used by TV manufacturers for advertising for a few minutes, after using H25 standard compression, the size is about 00.

Input the H25 video to FFPG,Decompress each frame of image data in JPG format,And then send it to the Gödel system,Of course, there will be a loss of image quality in the middle,But Wang Yinan can't care so much about it.,

He held his breath and kept his eyes glued to the screen.