Meng Fanqi decided on two directions for entrepreneurship, facial recognition and medical AI, with a sense of urgency between the two.

Facial recognition is a technology that has been used for a long time and is relatively mature in all aspects, but the previous methods were relatively traditional and outdated.

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

Medical AI is still in the early stages, and the most troublesome issue is the ethical issues related to medical data and patient privacy.

There are significant obstacles at the most basic level of data issues, and the procedures are cumbersome in all aspects.

Although the Shanghai Public Health Center took the initiative to contact him, it is unlikely that this matter will progress too quickly, and it needs to be carefully planned.

What needs to be dealt with first is the facial recognition algorithm, and since he has decided to start a business, he naturally needs to consider it from a commercial perspective, rather than the previous academic perspective.

Meng Fanqi understands the most advanced facial recognition algorithms of this period, such as Facebook's DeepFace, which was originally based on Alex's network for feature extraction, added segmented affine transformation, and used 3D facial modeling to reproduce facial features and align facial elements.

In Meng Fanqi's view, this method is extremely cumbersome, with as many as one billion parameters, although its performance on a large human data set LFW is 97.35%, close to the level of human performance.

However, for Meng Fanqi, it is almost certain that this performance can be further improved to above 99.6%.

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

From an academic perspective, this is not a big deal, as long as the world record is broken, it is naturally worth publishing research.

But in the industry, the thinking cannot be so simple.

In cases where the performance is almost the same, there are too many other factors to consider.

For example, the speed, commercial use, there are hard benchmarks for speed, and Meng Fanqi is very confident in this regard; and whether the algorithm's operators are relatively common? Some complex academic operations are not convenient for commercial use, and hardware devices may not support them, which may be a problem.

Other factors such as price, ease of use, the aesthetic level of the user interface, and even whether the promotional PPT is impressive or not, are likely to be one of the basis for laymen to make commercial judgments.

Find the original at " pa𝑤read.com ".

Therefore, Meng Fanqi believes that in the field of facial recognition, his breakthrough in just these two areas is a significant advantage, but it is not enough to establish an absolute advantage.

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

Meng Fanqi plans to build a strong technological barrier in this field, at least to make all other tech giants retreat for several months, or even close to a year.

Is facial recognition too simple now? Can it be achieved with old methods at 96-97?

Let me give you some intensity and see if you can handle it!

Meng Fanqi's strategy is based on his first public paper, Generative Adversarial Technology.

He plans to make some targeted adjustments to the adversarial generative network based on the residual network, and use the largest facial image data in the industry to train them.

The ultimate goal is to generate lifelike, but actually non-existent, facial images.

After this generation model is successfully trained, Meng Fanqi can use it to launch targeted challenges to the advanced facial recognition algorithms on the market.

Many of these facial algorithms on the market are based on traditional feature methods, and even the just-mentioned DeepFace has not been released.

Originally, they can only achieve a level of 94-95 at most, which is far from the 99.6 that Meng Fanqi can achieve.

On this basis, they also completely lack the ability to distinguish generated false images.

Meng Fanqi can freely use various false facial images to deceive these algorithms, and even generate corresponding facial images for specific faces, and deceive various security products based on these algorithms.

Completely shake the commercial value of the other party from the most fundamental issue of security.

Imagine, since there is already the ability to generate such arbitrarily fictional faces on the market, and Facebook's facial recognition technology has no countermeasures at all, it cannot distinguish at all.

This brings huge risks, and no one is sure what kind of product is being identified and allowed to pass through.

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

In this case, all first parties, especially government agencies that value security, will make the most sensible choice.

As the algorithm designer, Meng Fanqi is of course very clear about the problems and loopholes in such a generation strategy, and what regularities in the generated images cannot be discovered by humans.

Meng Fanqi's facial recognition algorithm will simultaneously have the first breakthrough in human-level accuracy, detection speed tens of times faster than the world's leading algorithms, and the unique and unparalleled ability to detect forgeries at present.

At the same time, the Facebook DeepFace team, who knows nothing about Meng Fanqi's new plan, is collectively studying Meng Fanqi's paper and code, completely unaware of what they will encounter."We are doing pioneering work in applying deep learning to facial recognition for the first time, and the scale of the data used is as high as millions. If so many algorithm components are replaced at this time, will it delay too long?" Among the DeepFace team, Yang Ming is the only Chinese, and he is slightly worried about this.

"Yang, now Meng's residual network has swept the entire AI industry. If we still use last year's 8-layer network, can it really be called the first work to apply deep learning to facial recognition?" Mrs. Wolf, as her name suggests, is very aggressive in her work.

In her view, Meng Fanqi has made a revolutionary breakthrough in the core of deep learning, the network structure itself.

If this new technology is not adopted, then the articles or code released by oneself are just a flash in the pan. A few months later, there will definitely be versions based on Meng Fanqi's residual technology everywhere.

Since I have realized my shortcomings, I must correct them, not be afraid of trouble, and not be afraid of not having enough time.

The open source release of the residual network was 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 hard in this direction for nearly half a year. Now it's just a matter of quickly iterating the final version of the experiment by replacing some components, which won't delay too long.

Can't others easily surpass the accumulation of so much technology for so long?

"Yang, you don't have to worry. Our main steps are detection -> correction -> re-expression -> classification verification. The later steps are already quite mature, it's just that now we have a better method of feature extraction."

Tegoman also comforted Yang Ming, knowing that this new member who joined Facebook urgently needs some achievements. "After changing the method, we can do better!"