Meng Fanqi decided to start a business in two directions, face recognition and medical AI, of which there are still priorities.

Face recognition is a technology that has been used for a long time, and it is relatively mature in all aspects, but the previous methods are more traditional and backward.

Once Meng Fanqi makes some breakthroughs, he can quickly cut into the battlefield, start harvesting, and make a quick profit.

Medical AI is still in its early stages, and the most troubling issues are the ethical aspects of medical data and patient privacy.

At the most basic level of data issues, there are quite a few obstacles, and the procedures and regulations in all aspects are cumbersome.

Although the Shanghai Public Health Center has taken the initiative to contact him, I am afraid that this aspect of the matter will not be advanced too quickly, and it needs to be done slowly.

The first thing to deal with is the face recognition algorithm, and since you have decided to start a business, you should naturally consider it from a business perspective, not from an academic perspective.

Meng Fanqi understands 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, adding segmented affine transformation, and using 3D face modeling to reproduce facial features and align facial features.

Facebook's method in 14 years is the foundation work of face recognition algorithms in the era of deep learning, and it has a strong influence.

However, in Meng's view, this method is extremely bloated, with hundreds of millions of parameters, although the performance on a large human dataset LFW is 97.35%, which is close to the human level.

But for Meng Fanqi, continuing to improve this performance to 99.6% is a sure thing.

However, it is clear from the data that the remaining room for improvement in this indicator is actually very small, and it cannot widen the gap very significantly.

It doesn't matter if you think about this from an academic point of view, as long as you break the world record, it is naturally a research worthy of publication.

But in the industrial world, thinking can't be so simple.

There are too many other factors to consider when performance is similar.

For example, the speed is fast and slow, commercial use, and there are hard indicators for speed, which Meng Fanqi is very confident in; For example, are algorithm operators more common? Some complex academic operations are not convenient for commercial use, and the hardware equipment may not support them, which may cause problems.

Other factors such as price, difficulty of use, aesthetics of the user interface, and even whether the PPT advertised is bluffed or not are likely to become one of the bases for laymen to make business judgments.

Therefore, Meng Fanqi feels that on the issue of human face, which has been relatively mature, simply his own two or more technological breakthroughs are only a big advantage, and it is not enough to establish an absolute advantage.

Since it is the first shot of entrepreneurship, it must not only succeed, but also win big.

Meng plans to build a strong enough technical barrier in this area to make all the other tech giants retreat for at least a few months, if not more than a year.

Is face recognition too simple now? The old way you can do 96-97?

Dude give you some intensity and see if you can stand it!

Meng's strategy is based on one of his first published papers, generative adversarial techniques.

He plans to make some targeted adjustments to the adversarial generative networks based on residual networks, and train them with some of the largest face image data in the industry.

The ultimate goal is to generate images of faces that look lifelike, but don't actually exist.

After this generative model is successfully trained, Meng Fanqi can use it to launch targeted challenges against the world's advanced facial recognition algorithms.

Many of these face algorithms on the market are based on traditional feature methods, and even DeepFace, which Meng Fanqi recalled just now, has not yet been released.

Originally, they were only at the level of about 94-95 at most, which was a lot worse than the 99.6 that Meng Fanqi could do.

On top of that, they don't have the ability to discern generative fake images.

Meng Fanqi can use all kinds of fake face pictures to deceive these algorithms at will, and even generate corresponding face images for some specific faces, and deceive various security products based on these algorithms.

Directly from the most fundamental issue of security, the commercial value of the other party will be completely shaken.

Imagine, since there is already such an algorithm ability to generate fictional faces on the market, and Facebook's face recognition technology has no countermeasures and is completely indistinguishable.

This brings a huge hidden danger, the truth is not distinguished, the product identification is successfully released, no one is sure what the hell is going on.

At the same time, the recognition accuracy and recognition speed of these products are far inferior to Meng Fanqi's technical products.

In such a situation, all clients, especially security-conscious government agencies, will make the wisest choice.

As the designer of the algorithm, Meng Fanqi is of course very aware of the problems and loopholes of such a generation strategy, and the generated images have laws that humans cannot discover.

Meng Fanqi's face recognition algorithm will also have the accuracy of breaking through the human level for the first time, dozens of times the detection speed of the current world-class algorithm, and the unique and unique semicolon forgery detection ability.

Meanwhile, Facebook's DeepFace team, which knows nothing about Meng's new plans, is collectively working on Meng's paper and code, with no idea what will happen to them.

"What we're doing is the first groundbreaking work to use deep learning for facial recognition, and it's going to take a million tons of data, so wouldn't it take too long to replace so many algorithm components at this time?" Yang Ming, the only Chinese in the DeepFace quartet, is a little worried.

"Yang, now that 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 a revolutionary breakthrough in the core of deep learning, the network structure itself.

If you don't adopt this new technology, then the articles or codes you publish will be short-lived, and in a few months, you will definitely be full of versions based on Meng Fanqi's residual technology.

Now that you have realized your own shortcomings, you must correct them, and you can't be afraid of trouble, nor can you be afraid of not having enough time.

The open source release of the residual network was a few days ago, and everyone was on the same starting line.

There is nothing to worry about.

The DeepFace team has been working hard in this direction for more than half a year, and now they are just replacing some components and quickly iterating on the final version of the experiment, which will not be too long.

With such a long period of technology accumulation, will others catch up and surpass at will?

"Yang, you don't have to worry, our main steps are detection-> correction-> re-expression-> classification verification, and the next few steps are quite mature, but now there are better feature extraction methods."

Tagman also comforted Yang Ming, he knew that this new young man who joined Facebook urgently needed some results, "After changing the method, we can do better!" ”