Professor Wang Shiguang of Machine Vision: Where is the Opportunity to Face the AI?

(Original title: Well-known professor of machine vision, Shiguang: Where is the opportunity for human beings in the face of AI?)

On May 21st, at the MTA Tianmo Music Festival, where music, technology and art were combined, Professor Shan Shiguang of the Chinese Academy of Sciences Institute of Entrepreneurship in the field of CV began to share the characteristics of this round of artificial intelligence from the perspective of computer vision. And some progress he has made after starting his own business.

First of all, Prof. Shiguang Shi listed some examples of our application of computer vision technology from the factual level. The computer has also begun to perform the task of “seeing pictures and speak” as we were when we were young. This is the autograph technique. The most typical application may be auto-driving with the participation of many companies nowadays. In addition, in the past five years, the error rate of computer classification and recognition of objects has basically been increasing at a rate that has fallen by half every year.

The advancement of computer vision technology has benefited from the promotion of three major engines: 1. The increase in computing power brought about by the popularity of GPUs; 2. The use of big data; 3. The revival of deep learning algorithms.

For the discussion of artificial intelligence, the always fascinating topic includes the comparison between humans and intelligence. Professor Shi Guangguang also shared his understanding of this type of problem.

He quoted the view of Mr. Li Kaifu on the occasion of public disclosure: 10 years later, artificial intelligence will replace many occupations in the world, and 50% of jobs may be replaced, including translators, reporters and assistants, including security guards, drivers, sales, etc. .

Perhaps there is a view that "the benevolent sees the benevolence and the wise sees the wisdom", then exactly how to judge what careers will be replaced will be the first to answer AI more easily in which areas beyond humans, and in which places have yet to be broken.

Professor Shiguang Shi introduced that AI is good at inductive learning through large amounts of data. In addition to induction, human learning also includes deductive reasoning, but it requires a part of inference, and there is currently no way to solve deep learning. Enhanced learning through self-judgment is also currently not possible with machines. Therefore, the AI ​​at this stage is more suitable for areas where data collection, acquisition, and labeling are convenient, including computer vision and voice recognition.

It can be seen that there are two major categories of areas where AI easily transcends humans: the first category is search problems or retrieval problems in a large amount of space; the second category is areas where skills are learned through experience, such as automatic driving and medical reading.

So where is the opportunity for humans? Professor Shan Shiguang believes that despite the fact that human beings do not make as rapid progress as cognitive ability, the human brain has logical reasoning ability in addition to data learning ability. Humans can actively design algorithm models for themselves, and can also actively collect data. Generic AI has not shown any signs yet. The current AI is targeted at specific areas.

Finally, Professor Shan Shiguang shared some of his own personal development. He founded China Topvision last year to provide users and customers of all industries with the ability to produce their own AI engines based on private data. At present, they have done face recognition for Huawei in mobile phones, and also cooperated with customers such as China Mobile and Ping An.

The following is the speech of Shan Shiguang:

Hello everyone! I'm Shan Shiguang from the Institute of Computing, Chinese Academy of Sciences. About two or three years ago, we entered a new round of artificial intelligence, and we called it the third wave of artificial intelligence. From the perspective of computer vision, that is, we hope that machines can look at the world like people, we can explore the characteristics of this round of artificial intelligence.

First of all, just to give a few examples, computer vision, that is to say, after the machine has a camera, what can it do? For example, the most typical example, autopilot, or car-assisted driving, especially with Tesla as the representative of the autopilot, or assisted driving, has been able to achieve on the road to the car, pedestrians, lane line Such as the detection and identification of such objects. At the same time, using the detected cars and people can help us drive.

Example 2: From the perspective of computer vision algorithms, in the past 3 or 4 years, we can clearly see that from 2012 to 2015, we let the computer classify correctly what it sees On such a problem, the error rate is basically increasing at a rate that is reduced by half every year.

Example 3: Autograph Technique

We can imagine that giving everyone a photo, let the machine automatically describe, or write a paragraph to describe what is in this picture. For example, if there is a photo, the machine can automatically generate a sentence to describe it in an open market, and there are many people shopping (vegetable market). This is similar to the task of reading and writing composition when we were young. This is also a very important task for computer vision.

In the past one or two years, everyone has also been increasing their number of faces every year. I believe that in the future, the number of times we will brush our face each year may increase to 10 next year, and the future will be hundreds or even more. We use such a system to brush your ID card to determine if you are the legal holder of the ID card.

With the advancement of such computer vision technologies, there are three major engines at work: 1. Very powerful computational capabilities, as we have already seen, especially the popularity of GPUs, allowing us to train very complex algorithms. 2, big data. Face recognition system, Google adopted 8 million people 200 million photos to train their deep learning model. At this point, it is impossible for anyone on earth to see so many people in their lifetime to train the face recognition algorithms in their brains. Our system can determine if he is the legal holder by swiping an ID card.

From an algorithmic point of view, it is the technology of deep learning. Deep learning is not a new invention, but a renaissance. Like the Renaissance, it is, to a large extent, a complex set of past history. Neural networks and deep learning are further extensions of multilayer neural networks that were popular in the late 1980s. When it was combined with big data, supercomputing and marriage, its power came into play, so that today's technological progress has been made.

In the past two or three years, there are many AI problems and tasks. The computer has gradually surpassed our human intelligence. This point is also a historical necessity. In fact, it has already occurred in many aspects. For example, we just saw our ID card to determine if it was your task. At the moment, computers can do it when 10,000 people try to impersonate you. We have a 95% probability that they can be correctly identified.

In what areas can AI surpass humanity? The algorithm that artificial intelligence mainly depends on is deep learning. The problem with deep learning is that I call it "data fertile," and it's good data. Good data fertility means that we have a lot of data for induction learning.

In addition to inductive learning, we also have a kind of learning called deductive reasoning or deductive learning. For example, look at the original Euclidean geometry, which is derived through reasoning. At present, deep learning is suitable only for learning from data. It is more suitable for data acquisition, acquisition, marking more convenient areas. For example, now doing computer vision, speech recognition, or more and more Internet-enabled areas make it easier for us to collect data. But it needs reasoning. At present, there is no way to solve deep learning.

Another important AI event took place last year. The Alpha dog defeated the Go champion. Deep learning in Alpha Dogs has achieved 80% of learning. There is actually another technique called reinforcement learning. It is suitable for the field that can automatically judge right and wrong, but it is not suitable for solving the problem of computer vision recognition. Machines can't judge by themselves, so it's very difficult to enhance learning through the accumulation of right and wrong data. If it is completely given to the machine, let it enhance itself, the current algorithm will lead to it biased, may get into the devil, learn silly.

Recently, Mr. Li Kaifu has introduced on many occasions. He believes that after 10 years, artificial intelligence will replace many occupations in the world, and 50% of the jobs may be replaced, including translators, reporters and assistants, including security guards, drivers, and sales. Wait. The benevolent sees benevolence and the wise see wisdom. Many experts do not recognize that all occupations are so easily replaced by our AI. But there are indeed many industries, and now more and more may be threatened by AI, such as Security.

This is where we are in a unit. They now use our face recognition technology to make access control. The entire group has more than 10,000 people. He can open doors and attendance (face recognition technology) for every person who comes to work. This system is also just running, and we also believe that when this system turns all its employees into mature people, it will certainly be much better than our human security. A good security guard can recognize 1,2000 people, but it is still difficult for a company with more than 10,000 people.

What areas will gradually be overtaken by AI? One category is the search problem of a huge amount of space, and the other is retrieval of categories, such as the retrieval of images. This is a piece of cake for the machine, and it is not so easy for us humans. Then there is the field of experience and skill dependence, which is the so-called knowledgeable, and through experience learning, acquired skills areas may gradually be replaced by AI, such as face recognition, object recognition, or autopilot, This is also an empirical issue, such as the medical map.

Our artificial intelligence can be combined with hundreds of top doctors. By studying these films, we can surpass many experienced doctors. Customer service quizzes are often semi-repetitive or fully repetitive. Therefore, it is entirely possible for artificial intelligence to learn such skills from historical experience.

Everyone will also ask such a question. Do you need to know how the human brain works beyond human intelligence? Can we be able to make a surpassing human ability algorithm? In fact, how our human brain works is still a very mysterious thing. It is also a subject worthy of study. The good news is that we don't actually need brains. If we can only stick to a brain-like line, we can't go beyond people.

Our current AI can be simply summed up as an algorithm, or a model, plus a method of data. Such a methodology makes our machine learn more about humans from a large amount of data than we can see from humans. It cannot be understood, but it will be superior to human representation and classification methods.

For example, Go, because of the appearance of alpha dogs, our Go experts and Go players have already begun to break through some of the previous thinking frameworks and learn from Alpha dogs. It also played chess that was not considered good in the past, but found that such moves were better.

So, is it like people? It is not a good or bad mark for this algorithm. For example, in face recognition systems, we now have no idea what kind of features a machine can learn through such a large amount of learning and can do better than others. This point has gone beyond what we humans can understand.

Where are the human opportunities? Human intelligence, in addition to algorithms, we have an algorithm for the brain, in addition to data learning, we also have logical reasoning. Compared to machines, our algorithms and models can be designed by ourselves. A very important feature is that our data is collected on our own initiative, not as passive as current machine learning algorithms, what data you give it, what data it learns.

We humans also have some very interesting features, such as our visual intelligence, and sometimes our mistakes are also very important parts of our intelligence. In the picture on the left, we can see that the brightness of the block and the block is the same or not the same? I believe no one really can see the brightness of these two blocks is the same. If you think that the colors of the two pieces are indeed the same, I believe your brain may have problems and need to see a doctor.

The color of the above block and the color of this block are exactly the same, but we will not be aware of such a correct result. In fact, we can think of the world as our own imagination. However, this kind of imagination is very difficult for the current machine, so that the machine can judge these two problems, and it can also accurately determine the two answers.

Our rational measurement of AI's progress requires us to pay attention in many ways. We have seen a lot of progress, but they are all progress in specific areas, and general AI has not yet appeared. Perceiving ability is also changing with each passing day, but our cognitive ability has not made much progress. The so-called perception ability is the ability to see, the ability to listen, and so on. Another point is that our AI at this stage cannot be learned on our own, let alone on our own initiative.

This means that the current AI is domain, experience, data-dependent, and it can only be AI in a specific area. Where is the general AI army? Including academia also does not have very accurate answers.

It must be an era of Spring and Autumn and Warring States, and AI applications in all industries will flourish. However, the time when Daqin unifies the whole country is still far from coming. Many industries need the production capacity of their own AI engines.

I myself also founded a company last year. We call it Keshituo. We have a Chinese-Western name called Che. We set up such a platform to provide such services for each bank. Industry users and customers provide the ability to produce their own AI engines based on private data. We provide engine and empowerment capabilities for face recognition in Huawei mobile phones, including some large customers such as China Mobile and Ping An.

Summarize briefly. In the past few years, the improvement of perception has driven the AI ​​boom. It should be said that the traditional industry can use AI to have a very good upgrade opportunity, but the general AI still needs time. So, deep learning is in a sense, we think it needs infrastructure. This is also a very important goal for the establishment of Christie. Hopefully, we will be able to turn to the road of AI technology. Thank you!

燑br>

Wireless Charging Multifunctional Sterilizer

Phone Sterilizer,Mobile Phone Sterilizer Box,Wireless Charging Uv Sterilizer Box,Wireless Charging Multifunctional Sterilizer

wzc , https://www.dg-wzc.com