Shulou(Shulou.com)06/02 Report--

[introduction] from December 5 to 7, the China big data Technical Conference (BDTC 2019), sponsored by the Chinese computer Society, hosted by CCF big data expert Committee, and co-sponsored by CSDN and Zhongke Tianji, was solemnly held in Beijing Great Wall Hotel.

Top 100 + technical experts and 1000 + big data practitioners gather here to focus on the development curve of big data's technology in the intelligent era with the theme of "big data driving Intelligence +", centering on the latest practice of the combination of big data and various social industries. carried on the in-depth interpretation and discussion.

At the main forum, Hu Yu, co-founder and rotating president of iFLYTEK, delivered a speech entitled "artificial Intelligence from Technology to Industrial Development path".

Hu Yu is a Ph. D. in signal and information processing at the University of Science and Technology of China, a professor-level senior engineer, a special allowance expert of the State Council, a member of the New Generation artificial Intelligence Strategy Advisory Committee of the Ministry of Science and Technology, and the Chief expert of the 863 key projects of the Ministry of Science and Technology; Executive Director of the National Engineering Laboratory for speech and language Information processing; head of the Application Planning Group of the Industry-University-Research Innovation Alliance of the Chinese Academy of Sciences Part-time professor and doctoral supervisor of the University of Science and Technology of China; Vice Chairman and member of the Chinese Society of artificial Intelligence; Executive Director of the Chinese Information Society; second-year student of Hupan University.

The core point of the speech: under the current development of artificial intelligence technology, deep neural network combined with big data has become the inevitable path of technology and industrial development. In the process of development, it is necessary to combine product innovation and business model innovation to obtain real application data. At the same time, achieve the landing of artificial intelligence applications, and continue to invest the commercial returns of the application into the core technology research. Realize the ripple effect of iterative and synchronous development of industry and technology.

The following is a transcript of the speech:

It gives me great pleasure to come to BDTC today to talk about "the path of artificial intelligence from technology to industry", especially the latest progress in the combination of artificial intelligence and big data.

Artificial intelligence continues to evolve

Artificial intelligence has been very hot in the past few years, but anyone who has something to do with artificial intelligence will say that they are engaged in artificial intelligence, just like a few years ago, everyone said that they were doing cloud computing, Internet and big data.

But in the past few years, we have been proposing to make a clear classification of artificial intelligence. We all know that the first computer in the world was invented to do things that humans are not good at or are not good at, such as computing. The characteristic of the computer is that it can calculate faster and save more. The mission of the first computer is to analyze such complex problems as the trajectory of shells flying in the air until 2017 and 2018. AlphaGo, which is very hot, defeated Li Shishi's ninth section, and so on, but did not make much breakthrough in the use of artificial intelligence, all using the computing power of artificial intelligence.

But why aren't humans and animals good at computational intelligence? Because human ancestors do not need to count if they want to survive on the African prairie, but they must look and listen, and they must know how to hunt and escape. An article on Nature specifically studied the neural circuits in which animal predators and prey perceive movement and found that they were the same. When people and animals competed on the African prairie, what they needed was perceptual intelligence and motor intelligence.

Machines have developed rapidly in this respect. Perceptual intelligence has made continuous breakthroughs in language and images in the past two years, and motion intelligence has been vigorously developed in the direction of industrial robots and service robots. So are these aspects the highest goal of the development of artificial intelligence? Actually, it's not. One of the important directions for human beings to distinguish animals is that human beings have cognitive intelligence. we put forward the direction of cognitive intelligence for the first time. I am glad that many research groups also regard perceptual intelligence and cognitive intelligence as the main research direction in the future.

Human cognitive revolution

Why is cognitive intelligence the research direction of computers? Let's take a look at a short story. Yuval Noah Harari, a very popular young Israeli historian, wrote the book "A brief History of Human beings", which points out that human history has experienced agricultural revolution, industrial revolution, scientific revolution and technological revolution, but the reason why human beings are different from animals is actually a very important genetic mutation. He found that there were many kinds of apes on earth, among which all our ancestors belonged to the descendants of Homo sapiens in Africa, while other apes were gradually defeated by Homo sapiens who came out of Africa and disappeared. Why do Homo sapiens out of Africa have so much ability? All ape-men can use fire and tools, but what are the characteristics of African ape-men?

Studies have shown that the greatest feature of African ape humans is that they have experienced genetic mutations, and the temporal lobe of their brain is transformed into a small tissue that produces language, and it is precisely because of the use of language that Homo sapiens in Africa can very well describe their surroundings, especially an ability that humans are born to be very good at, called "gossip", because gossip, human society can be established, because gossip. We have studied that more people can be organized. You can imagine that today's humanoid animals, such as gorillas, can only have 50 in a group, and if they exceed that, they will be separated, and they cannot organize larger social activities well without language.

Another very important advantage is that with language, human beings have dreams. Human beings can create religions and economies, so they also call this revolution the "human cognitive revolution." however, there is still a long way to go for computers to realize the cognitive revolution.

Cognitive intelligence includes language understanding, knowledge expression, logical reasoning and final complex decision-making. Cognitive intelligence is the biggest challenge of artificial intelligence. As the first step of cognitive intelligence, natural language processing has received a lot of attention and support from international leaders, but it is also an obstacle to the breakthrough of artificial intelligence.

The realization path of artificial Intelligence

Depth neural network

According to the research direction and way, I think artificial intelligence is mainly divided into three categories: the first is the deep neural network that everyone talks about now, which makes use of big data, supercomputing power and very complex model structure. for the first time, it has achieved some progress in perceptual intelligence and motion intelligence, especially in perceptual intelligence.

But here is a small question. Have you ever thought about where the data came from? The use of perceptual intelligence and big data intelligence to achieve artificial intelligence requires a large number of real big data, which is completely different from human learning. Human learning is a small sample learning and does not need too much big data, especially in voice and images.

Whole brain simulation

The second path for the realization of artificial intelligence is the "brain plan" put forward by the United States and Europe. These proponents have an idea that with the further development of supercomputers, they have a deeper understanding of the human brain. is it possible to simulate all the topologies and discharges in the human brain like a supercomputer simulating the global weather of a nuclear explosion? Because there are only 100 billion neurons in the human brain, it has been recently reported that some people have completely restored the nervous system of fruit flies and even mice. If the computer can simulate the discharge of all the brains one day, do we have an intelligence? This direction is put forward by brain science scientists, not to mention whether it can be achieved, but getting such a system does not seem to be the direction in which we now use electronic computers.

Intelligent dynamics

The third option is what we are doing. We hope that just as humans study the flight of birds, humans will eventually develop "aerodynamics" and build a better plane than birds, which do not have to eat and drink like birds, but faster than birds, because airplanes follow the basic theory of aerodynamics of bird feathers.

As we have a deeper understanding of the human brain and the topological structure and learning mechanism of the human brain, it is entirely possible for us to invent a kind of "intelligent dynamics". After learning the learning mechanism and structure of the human brain, it can be perfectly combined with the fast and stored characteristics of the computer to produce the possibility of strong artificial intelligence or even super artificial intelligence, which is our research direction. Many of the industrial applications discussed later are based on big data's artificial intelligence.

Ripple effect promotes A.I. Development from technology to industry

Since the launch of the Deep Learning Network in 2006, we have achieved great success in promoting the development of the industry, but there is a key point that determines our success or failure in big data's artificial intelligence. Training big data artificial intelligence needs a lot of real big data, and in many cases the real big data can only come from real applications, so the question comes: when your artificial intelligence training is not good enough, where do you get the real big data? What kind of people are going to use it? Here we put forward a very important idea, and put it into practice, and achieved good results, we call it the "ripple effect". What does this mean?

An overview of iFLYTEK big data's abilities

When the artificial intelligence algorithm is used by the industry and the public, the algorithm does not extend to everyone in an instant. There are always some people who contact first, and some people who contact later. These people who come into contact first will continue to contribute data to the artificial intelligence algorithm. Using these data to continuously improve the performance of artificial intelligence systems, one day, just like water ripples spread on the water surface, the water ripples affect the water surface in a larger area. But the error rate is relatively high, with the system getting better and better, there is always a moment when the water ripple is close to zero when it has not spread to the water surface, and the correct rate of artificial intelligence has reached a very high height, ensuring the use of a large number of personnel. so it is very important for us to talk about deep neural network and big data, and it is very important to combine application products with artificial intelligence.

IFLYTEK's big data ability can be divided into four levels, the first is big data resources, that is, the real application data we obtain through various industries and businesses, then the big data ability platform, and then the business big data. In all businesses, there is such big data in education, medical care, political law, and smart cities, and the top layer is called big data business. That is, in advertising and recommendation, in credit and other aspects of the use of big data, can become a virtual business to implement.

IFLYTEK has a set of Odeon platform, which forms a closed loop for big data's aspect, including data cleaning, data mining and data labeling system, from collection, storage, calculation and service. On the basis of big data's artificial intelligence, the iFLYTEK Superbrain Project was launched in 2004, hoping to realize the transition from "listening and speaking" to "understanding and thinking".

In the new round of neural network-based process, we found that there are two layers of amplifier effects. The first layer, the current artificial intelligence not only includes voice, image, natural language processing, they are all unified into one framework, that is, the deep neural network. Eat through the deep neural network, these core technologies will produce a very large platform improvement. On the basis of these core technologies, we can integrate with various industries, including consumers, education, health care, politics and law, and smart cities.

Two directions of the Development of AI Industry

In what two aspects will artificial intelligence be applied to human society in the future? We call one of these aspects "natural interactive intelligence". People imagine a beggar who can't read, but he can do things according to your requirements. You give him 5 pounds, and he can buy you a carton of cigarettes. He can communicate with you with sight, hearing and mouth. In the future, every machine will have the ability to interact naturally. Another aspect is that there are many specialties after the invention of characters, so one application field of artificial intelligence is to study the professional industry big data and expert systems, which can help many industries solve the scarcest resource problems in the industry. the scarcest resource is the expert.

Today, we share the application direction of iFLYTEK big data from three cases:

Application case 1:

What you hear is what you see, that is, the "iFLYTEK hear" system. With the help of the ripple effect, speech recognition was not easy to use at the beginning, and the speech recognition system was implemented on mobile phones for the first time in 2010. after 3-4 years of development, it has become the first Chinese language in the world to break through the threshold of practical technology. By 2015, more types of voice will be recognized and put on the screen.

Many artificial intelligence technologies based on big data are used, such as microphone array acoustic processing and speech processing fusion technology, the introduction of image processing factors, and the integration of more professional algorithms, natural language processing knowledge.

IFLYTEK won four recognition categories in the world's highest-level CHiME speech recognition competition. Under the interference of all kinds of noise, the recognition rate of the original system is very low. After processing, we improve the recognition rate to 100%. With the help of the ripple effect, the number of users continues to increase, and the error rate continues to decline. The real human-generated big data helps iFLYTEK speech recognition cross the threshold of practicality.

Eventually, iFLYTEK launched a range of consumer-oriented products, including the "hear" website, mobile APP, voice recorders, office books, and the world's only fully offline voice-to-text system.

Application case 2:

Intellectual science is a very important application point in the field of education. The use of artificial intelligence to assist teaching in the field of education is not only a hot spot in recent years, but also a rigid demand. Represented by education, medical care, security, government affairs and finance all need to use big data of the industry, expert system and machine learning depth neural network algorithm to establish a new generation of expert system.

This is applied to the image recognition technology. In the image field, iFLYTEK has done handwriting online recognition, offline recognition, and recognition for the content of photos taken in the natural environment. In the highest international level of ICPR MTWI test, iFLYTEK won three championships in word recognition, layout analysis and formula recognition. Xunfei also won the championship in a number of tests at the top meeting ICDAR 2019.

The ability of natural language processing is also needed in education. In natural language processing, there are neuron system DNN that simulates human brain, RNN-LITM that simulates sensory memory and short-term memory of human brain, and algorithms that simulate visual analysis layer of human brain.

IFLYTEK's latest award is to win the first prize in algebra, geometry and application fields in the 2018 International semantic Evaluation Mathematical Quiz. On the highest level of SQuAD in the world, the machine reading level of iFLYTEK surpasses the human level in two evaluation standards.

In the world's first physician qualification exam, iFLYTEK uses understanding robots to learn hospital textbooks and knowledge, and applies it to the medical qualification exam. Why can I take the medical qualification examination? Starting in 2013, the Ministry of Science and Technology and 863 plan to build a gaokao robot, hoping that the gaokao robot will take the gaokao after learning math, Chinese, geography and history, and eventually be admitted to a university and Tsinghua and Peking University. At first, everyone thought it was very difficult, but after more than five years of unremitting efforts, now the first phase of the project is over. I am the 863 chief expert of this project. In the end, I got 120 points in robot mathematics and 90 points in Chinese. Geography and history scored 40 or 50. Using cognitive skills, intellectual medical assistants scored 456 out of 600 in the medical qualification exam, while they passed with a score of 360.

What is the use of core technology in the field of education? We can see that they can fully automate the marking of the test paper, first scan the test paper and the contents of the students' answers, then identify them, and then use the artificial intelligence algorithm according to the recognition results, combined with the experience and knowledge of the original teacher's grade. finally, a system that can automatically generate examination homework results is formed.

What else can big data's artificial intelligence system do? China's dream of education for thousands of years is to teach students in accordance with their aptitude. Now, according to the results of automatic correction, that is, the knowledge points of big data and knowledge graph in the background, the artificial intelligence system can judge which knowledge points students do not master enough, and generate the required homework according to these knowledge points. not all students are stereotyped exercises. After using this system, the homework burden of students has been reduced by 30%, and students' interest in learning has been greatly improved.

In the whole education of big data, iFLYTEK now promotes artificial intelligence and big data's platform in many aspects of teaching, learning, examination, evaluation, and management, centering on the needs of teachers, parents and students. I believe that in a few years, we will see that more and more students in China improve their abilities through personalized and targeted learning, and students will abandon the sea homework skills, but make use of the students' own differences to complete this work. We believe that such an opportunity will certainly transform the current situation of education and give China a better future.

AI+ big data is the inevitable choice of scientific and technological innovative companies

To sum up, the combination of big data and artificial intelligence is an indispensable path for scientific and technological innovative companies. I think iFLYTEK is not a mobile Internet company, but a technological innovative company, just like Facebook, Amazon and Google in the United States, because when you do the core technology, you have to make sure that the technology is the most advanced in the world. In the past 20 years, many of iFLYTEK's technologies, including voice technology and artificial intelligence technology, have achieved good results in the highest international competition. Some even break through the threshold of Chinese in the world.

Second, as an enterprise, the necessary condition is to be able to find the right product innovation and business model innovation, to be able to make huge profits.

Third, the profits should be continuously invested in core technological innovation.

Finally, iFLYTEK has an advantage when it comes to doing Chinese, because language protects us. Now, with the change of the international situation, we are forced to do all the language processing in the world, and now we are on this road. I believe that in another 3-4 years, we will be able to achieve a very good core technological breakthrough in nearly 50 languages around the world.

Just like the aero-engine in the mechanical age, artificial intelligence is the aero-engine in the information age, and we are confident that artificial intelligence will do its best.

In this process, iFLYTEK's artificial intelligence and big data business will be closely integrated, and the big data trend report released by BDTC also shows that artificial intelligence and big data will be closely inseparable at least now and in the next 10 years.

Finally, I believe that iFLYTEK will certainly be able to achieve better results, but also be able to better counter the achievements of Xunfei in big data.

QroomA:

Q: from technology to leading consumers, you were originally in charge of spending money, and then you were in charge of making money, and technology must not be lost in this process, so how do you grasp the change of roles and stick to your original mind?

A: I think there are several factors of transformation. First, I found that doing research is similar to doing industry. There is a very important ability to do research, that is, the ability to predict the future and what areas of research to do. This requires a very thorough understanding of the future development of this field and own ideas, so that we can keep up with the pace of history and not be thrown away by it.

The same is true in the industry. When doing a specific thing and direction, the most important thing is to think about whether the strategic direction is right or not. The strategic direction mainly comes from your pre-judgment of the industrial trend. As I just mentioned, why do we divide artificial intelligence into perceptual intelligence and cognitive intelligence, why do we use big data's artificial intelligence method, and why do we propose our "intelligent dynamics"? it is to make insights into the development of many things, industrial development and industrial laws in the future, and at the same time adjust the direction to deal with it. I think this is the same part.

Second, there are differences. When we do education, every teacher wants to invent a technology that can be used in 1000 places. The more, the better. However, when you work in the industry, you will find that you can't do all 10 directions just because your technology can be used in 10 directions. When you have a technology, you should not only judge whether there are business model innovation and product innovation in addition to technological innovation, but also use your own technology or the technology that you did not invent. Concentrate all your energy in the direction that you can break through in the shortest time in order to realize the four steps of technological innovation that I am talking about. This is different.

Last but not least, I think it is necessary to do research and industry, especially the core technology industry, that is, to be able to withstand loneliness, to be able to sharpen a sword every ten years, to be able to sit on the bench for ten years of cold, and to rise abruptly based on accumulated strength. Because this and the tuyere of the Internet, the tuyere of the mobile Internet is not the same concept, it is like mountain climbing, to slowly accumulate knowledge and move forward, and finally become the wave itself, not a surfing enterprise. As we said, in every wave of personal PC, what is it that really becomes the wave itself? Is the chip, is the technical software, is the basic cloud computing architecture, big data's architecture. Therefore, in the past two years, we must be able to endure loneliness, work hard and rise abruptly based on accumulated strength.

Guest profile:

Hu Yu, co-founder and rotating president of iFLYTEK. Ph. D. in Engineering, Professor level in signal and Information processing, University of Science and Technology of China; Special subsidy expert of the State Council; member of the New Generation artificial Intelligence Strategy Advisory Committee of the Ministry of Science and Technology; Chief expert of the 863 key projects of the Ministry of Science and Technology; Executive Director of the National Engineering Laboratory for speech and language Information processing; head of Application Planning Group of artificial Intelligence Industry-University-Research Innovation Alliance of the Chinese Academy of Sciences Part-time professor and doctoral supervisor of the University of Science and Technology of China; Vice Chairman and member of the Chinese Society of artificial Intelligence; Executive Director of the Chinese Information Society; second-year student of Hupan University.

Https://www.toutiao.com/i6770515204187881991/

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.