Meng Fanqi decided to start a business in two directions, facial recognition and medical AI, and there are still priorities.
Face recognition is a technology that has been used for a long time and is relatively mature in all aspects, but the previous methods are more traditional and backward.
Once Meng Fanqi makes some breakthroughs, he can quickly enter the battlefield, start harvesting, and make quick profits.
Medical AI is still in its early stages, and the most troublesome issue is the ethical issues surrounding medical data and patient privacy.
When it comes to data issues at the most basic level, there are considerable obstacles, and the procedures and regulations in all aspects are cumbersome.
Although Shanghai Public Health Center has taken the initiative to contact me, I am afraid that this matter will not advance too fast and needs to be taken care of slowly.
What should be dealt with first is the face recognition algorithm. Since you have decided to start a business, you naturally have to consider it from a business perspective rather than the previous academic perspective.
Meng Fanqi understood the most advanced face recognition algorithms of this period, such as Facebook's DeepFace, which was originally based on the Alix network for feature extraction, added piecewise affine transformation, and used 3D face modeling to reproduce facial features. , to align facial features.
This method adopted by Facebook in 2014 is the foundation of the face recognition algorithm in the deep learning era and has a strong influence.
However, in Meng Fanqi's view, this method is extremely bloated, with hundreds of millions of parameters. Although the performance on a large human data set LFW is 97.35%, close to human levels.
But for Meng Fanqi, it is a sure thing to continue to improve this performance to 99.6% and above.
However, it can be clearly seen from the data that the remaining room for improvement of this indicator is actually very small, and it is impossible to significantly widen the gap.
It doesn't matter if you think about this issue from an academic perspective. As long as it breaks the world record, it is research worth publishing.
But in the industrial world, thinking cannot be so simple.
When performance is almost the same, there are too many other factors to consider.
For example, speed, commercial use, all have hard requirements for speed, which Meng Fanqi is very confident about; and for example, are the operators of the algorithm relatively common? Some complex academic operations are not convenient for commercial use, and hardware devices may not support them, which may cause problems.
Others, such as price, ease of use, the beauty of the user interface, and even whether the PPT is bluffing, are likely to become one of the basis for laypeople to make business judgments.
Therefore, Meng Fanqi feels that on the already mature issue of human faces, his two or more technological breakthroughs are only major advantages and are not enough to establish an absolute advantage.
Since it is the first shot in starting a business, it must not only succeed, but also win big.
Meng Fanqi plans to build strong enough technical barriers in this field to at least make all other technology giants retreat for a few months, or even close to more than a year.
Is current face recognition too simple? Can we achieve 96-97 the old way?
Brother, let me give you some intensity and see if you can stand it!
Meng Fanqi's strategy is based on one of his first published papers, generative confrontation technology.
He plans to make some targeted adjustments to the adversarial generation network based on the residual network and train them with some of the largest face image data in the industry.
Its ultimate goal is to generate face images that look lifelike but actually do not exist at all.
After this generative model is successfully trained, Meng Fanqi can use it to launch targeted challenges to the advanced face recognition algorithms in the world.
Many of these face algorithms on the market are based on traditional feature methods, and even DeepFace, which Meng Fanqi just recalled, has not yet been released.
Originally, they were only around 94-95 at most, which was far from the 99.6 that Meng Fanqi could achieve.
On this basis, they are completely incapable of identifying generative false images.
Meng Fanqi can freely use various false face images to deceive these algorithms, and can even generate corresponding face images for certain faces, and deceive various security inspection products based on these algorithms.
It completely shakes the commercial value of the other party directly from the most fundamental issue of security.
Just imagine, now that there is such an algorithm on the market that can generate fictitious faces at will, Facebook's face recognition technology has no countermeasures and is completely unable to distinguish.
This brings huge hidden dangers. It is impossible to tell the authenticity from the fake. If the product is successfully recognized and passed, no one is sure what the hell it is.
At the same time, the recognition accuracy and recognition speed of these products are far inferior to Meng Fanqi's technical products.
Under such circumstances, all parties, especially government agencies that pay attention to security, will make the wisest choice.
As an algorithm designer, Meng Fanqi is of course very aware of the problems and loopholes in such a generation strategy, and what rules the generated images have that humans cannot discover.
Meng Fanqi's face recognition algorithm will simultaneously have the accuracy that breaks through the human level for the first time, the detection speed that is dozens of times that of the current world-class algorithms, and the unique semicolon-free forgery detection capability.
At the same time, Facebook's DeepFace team, which knew nothing about Meng Fanqi's new plan, was collectively studying Meng Fanqi's papers and code, without any idea of what they would encounter.
"What we are doing is the first pioneering work to use deep learning for face recognition, and the scale of data used is as high as millions. If so many algorithm components are replaced at this time, will it take too long? ?" Yang Ming is the only Chinese in the DeepFace group of four, and he is a little worried about this.
"Yang, now Meng's residual network has swept the entire AI world. If we still use last year's 8-layer network, can this really be called the first work to apply deep learning to face recognition?" As her name suggests, Mrs. Waugh is very wolfish at work.
In his view, Meng Fanqi has made revolutionary breakthroughs in the core of deep learning, the network structure itself.
If you don't use this new technology, the articles or codes you publish will be just a flash in the pan. After a few months, versions based on Meng Fanqi's residual technology will be everywhere.
Now that you have realized your shortcomings, you must correct them without fear of trouble or lack of time.
The Residual Network was released as open source just a few days ago, and everyone is on the same starting line.
There is nothing to worry about.
The DeepFace team has been working in this direction for more than half a year. Now they are just replacing some components and quickly iterating the final version of the experiment. This will not take too long.
After such a long period of technological accumulation, is it possible that others will not be able to catch up at will?
"Yang, you don't have to worry. Our main steps are detection -> correction -> re-expression -> classification verification. The following steps are already quite mature, but now we have better feature extraction methods."
Tigerman also comforted Yang Ming. He knew that the young man who had just joined Facebook urgently needed some results. "After changing the method, we can do better!"
