"Okay." Yang Ming no longer insisted, but he always felt that something was not quite right, and his intuition told him that something bad was about to happen.
"Meng's new technology is really easy to use!" In the past few days of experimental iterations, Facebook's DeepFace team has made several major leaps.
First of all, the original bloated model size has been greatly alleviated. Because the operators are simple and easy to use, the speed is also much faster.
Looking at the training logs of the model, the performance of the algorithm is also improving all the way, and is likely to surpass their previous best results.
"I really don't know how he thought of it." Wolf let out a long sigh. Such a simple operation can solve a big problem that has plagued the entire field for several years.
"Actually, this situation is the most irritating. It would be great if he really came up with a very complicated theory and operation." Yang Ming could understand everyone's thoughts: "It would be great if his strategy was really very complicated." If it is complicated, we will be completely convinced and will not have any illusions or regrets.”
"But he got it done in such a concise and simple way. It feels like Columbus discovered the New World after sailing on the sea. There may be a lot of wisdom and hard work, but it always gives us a sense of self. It gives me the illusion that I might have a chance.”
Yang Ming's self-understanding is relatively clear, and he knows that this idea of "I can fuck me" is just a mirage, and he has not degenerated into a keyboard warrior after all.
Several researchers still quite recognized Meng Fanqi's research results, but they had no idea what kind of surprise gift Meng Fanqi had prepared for their latest algorithm that had not yet been released.
Face recognition can be said to be the direction with the most researchers, the widest range of applications, and the largest scale in the entire field of computer vision.
As an ancient topic, the task of face recognition has gone through the development from traditional pattern recognition to modern deep AI and has become quite mature.
But from beginning to end, it has been plagued by an unsolvable problem, which is the type of occlusion in the image.
Ordinary photos are not Photoshop layers. If an object is obscured, this part of the information is completely lost.
There is no technical means to restore the scene at that time, just like demosaicing, it is impossible, and the destroyed layers are irreversible.
But now, we have a new way to find another way to solve this very difficult problem, which is Meng Fanqi's generative adversarial network.
Through repeated confrontational generation of large-scale data, the generation network will have very terrifying image generation capabilities, and can completely generate content that is very close to reality to fill in the occluded parts, although the generated content may have nothing to do with the original content.
But it does look like that, and it doesn't affect people's understanding of the image.
What's more, if it has seen quite a few other photos of the target person in advance, the generated part is likely to be almost exactly the same as reality.
Moreover, generation is only the most basic aspect. By understanding the parameters in some dimensions of the hidden layer, people can even finely adjust some attributes of the image.
The editing of specific attributes of images is a more advanced application of generation technology.
For example, in the image, the emotions of the characters and the joy, anger, sorrow and joy of the facial expressions can all change accordingly as you adjust them.
The fake smile emoticon pack that became very popular later was to control the emotions of the face in the image, making the originally sad and angry person grin and laugh exaggeratedly.
Some people even modify the entire video continuously to make the whole scene look very funny and hilarious.
In addition to expressions and emotions, another thing that became more popular later was age adjustment. By modifying the parameters of certain specific dimensions, people can grow old overnight or rejuvenate.
These are just regular image edits, and only the face is edited.
There will be more high-end algorithm applications later, such as smart face changing, smart beautification, etc.
The former is very popular among many film and television creatives and ghosts, and everyone can 'act' in a three-minute blockbuster.
Anyone you want can make a cameo, Guan Gong and Qin Qiong are possible.
The latter is widely used on various live broadcast platforms. Throwing two slippers on the ground can automatically apply makeup on you.
"However, due to current limitations of hardware and other technical means, I have not yet been able to generate very high-resolution clear fake images, and it is a bit difficult to directly operate the entire video."
Meng Fanqi clearly knew what step he could take now. There was nothing he could do with high-definition pictures with 4k resolution, but ordinary resolutions of around 2, 300 or even 500 were almost enough.
For the face algorithm at the current stage, it is more than enough.
Moreover, Meng Fanqi now has Google Brain's researcher privileges. It is not a problem to use dozens of graphics cards. The speed of doing these experiments has increased dozens of times compared with before.
After the training of this model is completed, Meng Fanqi plans to build a website called [None of These People Exist] to raise the profile and make everyone pay attention to this face forgery generation technology.
Every time you open this website, a portrait of a lifelike but non-existent person will be randomly generated.
Meng Fanqi plans to use his relationship with Google and whiteness to promote this website. In this way, he will remind people that the original recognition algorithm must be able to distinguish whether the image is real or generated by artificial intelligence.
It can be said that it kills two birds with one stone. It not only promotes its own generative technology, but also raises a big problem that cannot be solved in a short time in the current direction of face recognition.
With the time difference of about a year, Meng Fanqi, who has been able to easily and properly handle this problem, will form an unbreakable technical barrier and quickly seize the domestic facial recognition market.
"But even if this matter is more important, it still has to be ranked behind the recommended advertising algorithm." Meng Fanqi frowned. If he wants to start a business, there will be a lot of trivial matters in the future.
The recommendation algorithm cannot be delayed any longer. It is currently its biggest cash cow and will also be the core source of start-up capital for many companies in the future.
I can't let the small lose the big. I will go to Google's Shanghai branch tomorrow to start the research and development of this part of the content.
Until the test results are successful, everything else has to give way.
Whether it is FaceGAN facial generation technology, establishing a company, or subsequent plans to establish [these people do not exist] website to promote this virtual generation technology, it must be ranked behind the recommendation algorithm.
Wait until these contents are completed, and then take the time to go to Heshanghai Health Center to discuss the issue of intelligent processing of medical images.
