In addition to the fact that the facial recognition method promoted by Meng Fanqi when he was in China has become very popular, his recent series of work in the field of medical imaging diagnostic analysis has actually been trialled in California circles.
How do you say that? [It was unanimously approved by hundreds of professors].
Especially the two top universities near the Bay Area, Stanford and Berkeley. Because they still persisted in research when artificial intelligence was in decline, they also cherished these late fruits.
Liu Yong, a professor of oncology at Stanford, was very surprised after using this series of algorithms. This computer had just learned to distinguish a cat a year or two ago, so why is it suddenly so powerful now?
Especially for more difficult diseases, it would take a few students of mine to distinguish them within ten or twenty minutes. This thing is really good, and they can give an answer in just a few seconds.
To this end, he had private conversations with Meng Fanqi two or three times, roughly understood the principles, and provided a lot of data to support his coordination.
Once, Professor Liu Yong asked Meng Fanqi: "Since artificial intelligence can make such accurate analysis and judgment on the types, areas and contours of lesions in pictures, can you also help write medical orders or text analysis?" "
Meng Fanqi was speechless after hearing this. Unexpectedly, Professor Liu accepted new things very quickly and was already daydreaming about it.
He had to tell Professor Liu the truth, not to mention the multi-modal combination of language and images, the language model itself is still a field that is in urgent need of breakthroughs.
I'm afraid it will take several years to realize the functions he requires.
Other medical professors at Stanford, such as Director Jeffrey and others, are very optimistic about the Meng Fanqi Alphafold project, especially Jeffrey. He has served as the chief investigator of more than 20 clinical drug studies, so he is very capable of understanding this kind of protein analysis. The value of ability.
Jeffrey is a big-brained, somewhat chubby middle-aged man with a red face who treats his students very enthusiastically.
After learning that Meng Fanqi's Alphafold project lacked high-quality protein data, he also actively provided assistance.
Generally speaking, the Stanford medical school is very enthusiastic about the breakthroughs made by its students. With the influence of these two top universities in California, Meng Fanqi's achievements are slowly radiating outward.
However, the preparation of a large amount of data cannot produce significant results in just a few days.
Even with the collective support of relevant professors in California, the speed of accumulating data is still much slower than Meng Fanqi thought.
The official start time of the Alphafold project will be delayed by at least one to two months.
Therefore, during the period before the start of school, Meng Fanqi's protein analysis plan had to enter a period of stagnation.
"Technology advances too fast, so we will encounter this kind of problem." Meng Fanqi was a little helpless, not to mention that many fields had not had time to accept and digest the ability improvements he had made.
The little data originally accumulated in these places is completely inadequate in the face of new technologies, and the quantity is too small.
Even if they can turn around in time and start accumulating and annotating high-quality data, it will definitely take some time.
The data cannot keep up. No matter how good Meng Fanqi's theory is, it is difficult to produce good enough results to be convincing.
"I have been working on application technology for several months. Now I have to wait a month or two for the data. It seems it is time to do some basic work to pave the way for the future."
After many months, Meng Fanqi was finally forced to calm down and have enough time to do some basic theoretical methods instead of rushing to make products or monetize.
After all, whether future technologies use AI to generate speech, images or text, the current technical theories still have many flaws and problems.
What he plans to develop recently is a very popular image generation technology, the basic components of Stable Diffusion and the diffusion principle.
This was the cornerstone of many great later generation techniques, and is perfect for preparation now.
Diffusion model is a relatively unclear term. Although few people later knew the principle, many people have heard this term many times.
From the works generated by AI drawing software, which defeated many human artists and won the digital art championship, to later, domestic and foreign platforms such as Midjourney, Imagen, and novelai blossomed everywhere.
More and more people have clicked on relevant websites and tried to let AI describe the pictures in their minds, or make local modifications and adjustments.
Some use pictures to create various mysterious spells to summon ancient gods, while others use pictures to create pictures and make all kinds of magical jokes.
In 2022, AI painting and AI-generated images have advanced several times in just a few months.
Every progress and breakthrough has brought visible improvements, far beyond human imagination.
Just around the end of 2022, everyone was still laughing at what AI drawing was, it was too ugly.
As a result, three months later, it was discovered that things did not seem to be that simple. AI began to produce a large number of various turbulent drawings, which attracted a considerable number of people's attention.
At that time, many people were still joking that although AI was not good at drawing, don't tell me, its grasp of the subject matter was still very good! If the level is not good enough, the subject matter will come up.
In another three or four months, by the end of 2022, no one will dispute the level and ability of AI painting.
This time, the main point of debate has changed again. It has become whether AI drawing is plagiarism, and who is better, AI level or artist.
Putting aside the unclear question of who is stronger, it can be seen from the content of the debate that the ability of AI painting has indeed improved very quickly.
"To be fair, when it comes to AI generation, the GAN generative method has made everyone take a detour."
Although Meng Fanqi's GAN generative method has been widely praised in the academic community and has established a high academic reputation and popularity, and FaceGAN's fake face generation effect is also quite amazing, in the end it was the diffusion model that really made AI mapping popular.
"The adversarial model generated by the GAN method is certainly amazing, but it is still very troublesome for two networks to learn against each other." Meng Fanqi thought about it for a while. The big problem now is data on the one hand, and computing equipment on the other. It's the other side.
Originally, I released the technology in advance, but the graphics card was already insufficient. At present, GAN is used to target a specific thing. For example, FaceGAN can only do faces, but it is very difficult to generate it directly from text.
The principle of the diffusion model is actually not difficult. It mainly adds noise to the photo, and then learns various features of the current image in the process. Then randomly generate a noise image that obeys Gaussian distribution, and then reduce the noise step by step until the expected image is generated.
Writing code is not that difficult, but if you write it into a paper and ponder the principles, the mathematical logic and derivation in it are enough to make Meng Fanqi drink a pot.
"In terms of mathematics... let's ask Han Ci and Dean Fu for help in the past two days. I haven't contacted them for a long time. Or you can ask Xinton and Li Feifei for advice. These two are my mentors. .” When it comes to mathematical problems, it’s natural to find math professionals.
Although Li Feifei and Hinton are not from mathematics background, Li Feifei has a background in physics and Hinton is a godfather in the field. Both of them must be very strong in this area.
After creating a new folder, Meng Fanqi was about to start working, but he vaguely felt that something was wrong. He seemed to have forgotten something.
It took him a while to remember that his unlucky roommate Don Juan was coming to the Stanford area today.
"In the past six months, I almost forgot that I am still a student."
Meng Fanqi, who has been working intensively on research and development for three or four months, feels as if more than a year has passed.
It just so happened that Don Juan was here, so I gave myself a few days off. It had been a few weeks since I arrived, but I had never been to the Stanford campus.
