Huawei: Live From Huawei Connect 2019 Day 1 Keynote

Good morning. I join you here in Shanghai at the Huawei Connect conference. The SKA is a revolutionary radio telescope that will help us to study and understand gravity at a fundamental level, and utilizing the SKA's unique capability to study hydrogen, the most common element in the universe, to uncover the secrets of what we call the cosmic dawn and the birth of stars and galaxies. Ambitious projects like the SKA can only happen when governments, scientists, engineers, and industry come together with a common goal. The SKA Organisation, which has been managing the design phase of the project, is supported by 13 countries and brings together hundreds, even thousands of engineers and scientists from institutions around the world. In the SKA, the sun truly never sets. China is strongly involved in the project and has actually been a key partner since its inception. Next year, we will witness the birth of the SKA Observatory, solidifying further the partnerships between the many nations involved and providing the stability that the SKA will need for the coming decades. The SKA will then become an intergovernmental organization, similar to CERN and ESO, and only the second in the world dedicated to astronomy. 'Be brave to explore' was the instruction from China's Vice Minister of Science and Technology, Zhang Guoqiang, when he signed the convention of the SKA Observatory in March. That is what we do. So what are we actually building? To achieve the levels of sensitivity and spatial resolution we need, and to cover the range of frequencies required, the SKA will use many connected antennas, known as arrays, which can act together. The SKA will have two such arrays with different technologies, one in South Africa and one in Australia, both in very remote regions of the world. These remote locations allow us to get away from artificially generated interference, which would pollute the extremely faint radiations that we wish to detect. In South Africa, we will be building close to 200 large dishes, and in Western Australia, there will be more than 130,000 smaller dipole antennas. But this is just the start. We are aiming to greatly expand both telescopes in the future, and all of this will be managed from our global headquarters at Jodrell Bank in the United Kingdom. You can see on the screen behind me what the telescopes will actually look like. But in fact, we already have hardware on the ground, and before official construction starts in 2021, we have built prototype instruments and deployed them on site. China has been leading the design of the SKA dishes and have built two prototypes, one of which is at the CETC-54 factory at Xiji County, and the second is currently being tested on site in South Africa.

When you build an enormous radio telescope, the aim is to gather as much information as possible from the universe, and that means dealing with a lot of data. The SKA antennas themselves will send a huge volume of data to the massive digital systems and supercomputers, one in each of the two host countries, where the data will be processed to allow astronomers to undertake their scientific investigations. This is a data-intensive process that will initially require around 50 petaflops of dedicated signal processing power, growing to 250 petaflops as our capability increases. In total, every year we will archive around 600 petabytes of data, dwarfing the amounts currently handled by platforms such as Facebook and Google. Major research infrastructures like the SKA have a long history of driving innovation and making an impact on wider society. I'm sure you are all familiar with the fact that CERN in France and Switzerland developed the World Wide Web. Maybe less known, but equally relevant, is the development of Wi-Fi that a large majority of you in this room are using probably right now. This was invented by radio astronomers in Australia who were trying to process signals coming from black holes. Of course, I didn't bring my crystal ball today, so I can't really predict what the spin-offs will be generated by the SKA, but there's an enormous potential for us to have a similar impact through developments in areas like data visualization, artificial intelligence, and machine learning. In fact, our partners here at the Shanghai Astronomical Observatory have already been working with Huawei in this area, applying machine learning techniques to astronomy problems such as pulsar searches and radio galaxy detection using simulated SKA images. They will run a demonstration here in a few minutes. AI's ability to speed up data analysis will be invaluable when the SKA is operational, when we're dealing with these enormous data volumes. To make the 600 petabytes of data accessible, we are making an alliance of regional data centers, analogous to the CERN tiered computing network. This will allow scientists from around the globe to explore the data in new ways and hopefully unearth new scientific discoveries. Our colleagues at the Shanghai Observatory are leading on this front, and they recently completed a prototype SKA data center, which will provide us with great insight into how SKA data could be stored and accessed in the future. The excitement and momentum within the SKA partnership is building as we approach the start of construction, and I look forward to the day in 2025 when scientists from around the world, including here in China, will get their first glimpse of the universe through the eyes of the SKA. Thank you very much.

Thank you, Mr. Xu, for sharing Huawei's computing strategy and AI cluster launch. Next, please welcome Professor Gao Wen, member of the Chinese Academy of Engineering, director of Pengcheng Laboratory, professor of Peking University, and an ACM Fellow, to give a speech on 'Pengcheng Cloud Brain: Powering an Open Innovation Platform for AI'.

Distinguished guests, friends, good morning. Just now, Ken gave an excellent speech. Indeed, connectivity is more important to our age. In this era, computing power that connectivity will enable us to do another lofty mission. In early days, as you know, computing was meant to resolve engineering problems, such as the trajectory of guiding missiles, ballistic missiles. So for those computing, you have a fixed formula and certain parameters as inputs, and the computers will resolve the problem by working the equation. That is some computing model with certainty, as Ken mentioned. So computing power is increasing. We are not satisfied with only asking the computer to resolve such equations or functions. We hope that computers can also think like human beings and recognize images, videos, understand speeches. Then the computing will be totally different from the computing of equations. A lot of statistics will be required, and a lot of data need to be fed into machine learning, like the millions or tens of millions of images for machine learning. As I mentioned, there's a lot of throughput of data and also a real demand for superfast computing in the case of data processing. And this is for engineering. But what will be the future scenario of computing? The future will be computing like our brains. There is a long way to go, but in order to reach our destination, from this moment, we need to think about how we can migrate from intelligent computing to a semi-brain computing. Next, I would like to share with you what we have done to make this happen, what we have been thinking about, and what platform and devices we have put together. We call it Pengcheng Cloud Brain. The main base is headquartered in Shenzhen. As you know, Shenzhen is at the forefront of the reform and opening-up drive. Last year, in Shenzhen, Guangdong province, set up a provincial lab with a target to develop into a national lab. In the US, as you know, there are a lot of national laboratories. Each of them has a specific orientation and tasks. In China, there weren't national labs, at least those in the real sense. Now China is richer and can afford certain basic research. Therefore, China is thinking about developing the national laboratories, platforms, and facilities that go in the US. The Pengcheng Lab was established in Shenzhen. It is expected to undertake the basic research that can honor the national needs for development and security, and so on. And those demands were addressed in a discrete manner in various universities and research institutes. Now we have a new platform formed by such national labs, and Pengcheng Lab is one of them. In the information field, it is an upcoming national laboratory. The information, of course, we will talk about communication, connectivity, as well as computing. And for computing, we position it for AI. For this moment in computing, it is very important for you to develop the Pengcheng Cloud Brain. With that, we hope to cover intelligent healthcare, intelligent transportation, like the traffic jam now we suffer so frequently. So we are going to address all kinds of problems facing the country's development. So we need a model, algorithm, and research platforms to make AI happen and helping us address such problems. In order to address such problems, we need to develop three levels of facilities. The first level is a very strong software platform. As mentioned, computing needs to happen on a platform. Therefore, the cloud brain needs the fundamental level of hardware, and on top of it, we need software that can support an open-source model. Working in the world, there are so many open-source software. For all the mainstream ones, we provide support, and we are also going to put together a more efficient platform, which can make the things in perfect, better. So this is an open and open-source platform for coexistence of multiple systems. On the top level, we are going to develop certain typical applications to empower the society with AI. And the fundamental level, what so where shall we choose and how much computing power shall we plan? These are the key questions we need to answer. We need to support the major applications for AI and train models for them. For example, to train models for facial recognition, you need very good model training. In certain enterprises, facial recognition has been well done. The secret is my data is used to train the models, so the accuracy can be boosted to 97, 98, 99 percent. So that is only one of the many areas. So some other areas have been tried but not very successfully yet, but we can try to train all sorts of models on our platform. How can we structure this platform of hardware? We typically buy NVIDIA and other foreign GPUs, and in China, we also hopefully will use domestic processors. So we have developed such a platform over the past year, but this platform has now generated our computing power of 100 P. But that is not enough for major tasks, and we are going to scale it by 10x, so that it can be 1000 P. So what is going to be the technology that enables this new platform? I'm happy to tell you that we are going to use the latest Huawei Ascend for AI computing. As I mentioned, this platform is scalable. It is not just going to be complete in one go. Depending on the task requirement, it will scale from 100, 250, to 500, and then to 1000 P. So it will scale gradually. Along the way, the evolution is going to be a dynamic process. It will go with the upgrading and updating of software and hardware and system planning. So we are going to also support the clustering of AI. So this is how we design our roadmap. It not only has very strong computing power but also a lot of possibilities for researchers, students, young scientists to use data. So we are going to collect enormous image libraries, which will be open to the researchers and students who don't have data. But if they have, they are welcome to load their data into such libraries. So such a platform is critical to future science. We can think of a lot of domains of application right now, but for some future unknown domains of science, like what is happening in space or in the logical structures, so all these areas of interest can be studied by researchers and students after our platform is open to them. Once we have the hardware layer, we need to make it convenient for users to use. So to increase the convenience, we have put together the engine of computing for the cloud brain, which has an open-source framework. You are using or you are adopting, as long as you have the source code and the computing framework, this open-source Octopus engine for computing will provide the maximum possible support. We are also going to organize a complete ecosystem, and end-to-end support can be provided to domestic researchers, which may have or lack something. So we are going to complete the offering to make the research sustainable for the long term. We call it Octopus. There are different systems, basic research, arms, software, training, architecture. So in this platform, in terms of how to manage the hardware...

Thank you, Mr. Xu, for sharing some practice on the exascale supercomputing system. Now let's invite Mr. Zheng Li, the President of Huawei Cloud BU, for his presentation: Crossing the Commercial Chasm, Building Inclusive AI.

Good morning. The topic of my presentation today is Crossing the Commercial Chasm, Building Inclusive AI. Please look at this previous slide, which is very interesting. We can see two curves on both sides of the slide. On the left side, we can see that before 2018, the AI field, the financial and financing volume of AI is spread across different industries and technical fields evenly. It's like spreading things evenly across the entire field. At that time, investment and financing reflect the characteristics of the passion in this field. However, after 2018, we can clearly see another trend. A lot of financing and investment is centered around a few specific industrial scenarios. Meanwhile, it is highly concentrated in a few key fields such as smart mobility, smart city, healthcare, finance, to name a few. So from this chart, we can see that technological passion has been retraced back to the rationality in commerce. To make a new technology inclusive, there are chasms for us to transcend. Mr. Kang, who mentioned in his presentation that the computing power on the cloud has entered into the brute force era, and I call it violence. Theoretical computing power increases by tenfold each year. Why is that? Because from the research perspective, in the past two decades, the research is getting more in-depth, and we can see a growing number of new algorithms and new explorations. For example, in this chart, we can see BERT, GPT, to name just a few. These new algorithms have increased a hundred times of computing power. There's another interesting discovery: that is, over the past three years, you can see the growth of computing power is positively correlated to the number of published papers of AI. In the future of commercial application, only if the algorithm is cost-effective enough can we see the wider application of AI in commercial scenarios. Therefore, we hope that AI computing power can be readily accessible and convenient as electric power we have today. We hope it can be convenient and readily available. Two years ago, also in Shanghai, I shared with you my opinion. I said that computing power will unleash the beauty of algorithms. Today, if you have access to a large amount of computing power, you can test your model and train your model faster than your competitors. At our last year's Huawei Connect, Mr. Xu shared with us his idea. He said that Huawei's capacity to unleash AI has two dimensions. One solution is AI towards the intelligence, the other is towards universities, organizations, research institutes, by which we call AI for All. We also shared with you our opinion. We believe that AI can change three scenarios of productivity. One is repetitive high-volume works in our world, the second is expert experience in real, and the third one is multiple domain collaboration scenarios. How can we empower industry upgrading in these three scenarios so as to achieve efficiency enhancements, expertise transmission, and go beyond the human limits? For the repetitive high-volume works in our world, we mean that in enterprise practice, we need to find those highly repetitive scenarios with very frequent intervals. To give an example, a famous express delivery company called SF Express. We work together with SF Express. Through highly accurate OCR, we can scan the information on delivery sheets so as to replace the purely manual inputs. Express persons can take pictures or take screenshots, and the OCR system will automatically recognize the information of sender and receiver on the sheet. The system can also be used to deal with complex backgrounds, including uneven lighting and unclear pictures, but also even deal with missing pictures. In this way, we can reduce the manual processing time for abnormal situations, increase management efficiency dramatically, and improve user experience. The overall management cost has been brought down by 25%. In the express delivery sector, violent sorting and picking has always been the source of complaints. This behavior will harm consumers' rights and interests and also lead to a large number of claims for damage. The user will have a very poor experience. In this situation, Huawei Cloud Eye intelligent analysis system can have real-time behavior analysis of monitoring video and automatically identify the improper behavior of the pickers, such as tossing, throwing, or even kicking. Now, across 150 operation fields, we have installed 30,000 cameras which have been connected with AI identification systems for improper operation. What about their results? We can see that the hourly rule-breaking operation rate has dropped by 45 percent. The next scenario is expert experience in real. It is easy to understand. It means the accumulation and industrial know-how is integrated in the scenario so that AI can reach the level of assistance to AI. Last year, we worked together with a third-party medical examination institution, which is a listed company called iKang. We worked together with the company, and for the first time, based on personal mythology and through deep learning technology, we have trained an AI-assisted cervical cancer screening model which is very accurate and highly effective. In the past year, the accuracy rate of negative results reading has been over 99%, and the detection rate of positive results reached 99.9 percent. This is the highest level of AI-assisted cervical cancer screening known in the public. Let me share with you another case. It takes a highly qualified cytopathology doctor to read a Pap smear on the microscope six minutes. However, AI only needs 36 seconds. The third scenario is multi-domain collaboration with multiple parameters and complex interdependence. This is typical in industrial management, urban management, as well as logging-based intelligent reservoir identification, knowledge mapping, NLP, and other artificial intelligence technologies. New thoughts, new ideas will be brought into all these sectors. In the past eight years, we have been working on new industries, and we keep saying that Huawei always honors its words through action. In the past year, our Huawei Cloud AI capability has covered over ten industries, including city management, industry, retail, finance, automotive, and family. It includes over 500 projects. Working together with our partners, we help them to embark on the road of intelligent upgrading. As I said for many times, Huawei itself is a global leading electronic manufacturing industry. We understand clearly what problems need to be solved in industrial scenarios. I also shared my ideas with my colleagues in Huawei. What is the idea of B2B? It means your data, your model. Actually, we can help you to solve the problems in your platform. So it is all oriented for our industry partners. The successful practice of these 500 projects, together with well over 100 AI scientists and mathematicians from Huawei, who have reviewed and summarized our practice in these projects, we have found that to implement industry AI projects, it takes not only the platform and enterprises but also joint efforts of many players. That's why we propose this role model. We can see four players are included in this model, including AI platform, ISV, business scenario, and divisive systems and process. So our demand comes from practice, and implementation also happens in the practice. You can see that successfully implemented AI projects need four keys, which is a clearly defined scenario, readily available computing power, continuously evolving service, as well as matching organization and talents. If I look at this model, combining with four keys that will affect the success of a project, we can understand where the commercial chasm of the commercial application of AI technology is. We need to transcend that chasm rather than fall into it. The first key factor is a clear business scenario. We need to understand the problem at its core. Is it about quality? Is it about lowering cost? Or is it about efficiency enhancements? Secondly, which is also important, to solve this problem, we need to have a clearly defined boundary. This boundary should be expressed and defined mathematically. It must be expressed and defined mathematically. Certainly, this scenario should be a closed loop that is predictable rather than an open and unpredictable one. Last but not least, we should be equipped with sufficient data and complete industry know-how in order to solve the problem. That is how to solve the problem related to the scenario we have clearly defined. Let me share with you two cases to explain what kind of scenario is relatively easy and relatively complex. Salient Hope is one of our clients. This scenario looks complex, but the scenario actually can be clearly defined because it has a clearly defined boundary. The problem itself is clear, and the boundary can be expressed and described with mathematical means. We also have sufficient data and know-how to solve the problem. The problem itself is about the color inspection of synthetic fiber. How to improve the quality inspection and also solving the problem of dyeing and missing color to realize commercial value? I would like to invite you to enjoy the next video clip to see how Salient Hope made the leapfrog in quality inspection from merely 100 meters to over ten thousand kilometers with the help of Huawei.

Just now, the children from Salient Hope showed us a scenario that is clearly defined, which can be solved by AI successfully. The next case is not as simple as it seems. The implementation process has been quite difficult. The case is about river restoration. The scenario seems to be simple, but the similar simple scenario is extremely complex. In the beginning, everybody has very high expectations, but during the implementation, we discovered that the scenario is far too complex. First, we need to identify dropping poles, floating litter, and other scenarios. To make things worse, we also need to recognize complex natural environments such as darkness, rainy days, stormy weather. These natural environments also intertwine with human behavior, which makes the level of difficulty even higher. In these situations, the scenario is very difficult to identify. Luckily, we have a very good and determined client. Walking together, actually, we have a very good team and a very good customer. After numerous adjustments, we have conquered difficulties one after another, and we have adjusted and optimized our algorithm continuously. In the end, the recognition rate improved from the initial 50% to 90%. So I shared with you these two cases to tell you how to find the appropriate scenario. It is not a subjective judgment, but rather it is based on the nature of AI. If the scenario is not appropriate, this would be a beginning of a very difficult journey. The second key is computing power readily available. Strong computing power. Mr. Kang, we talked a lot about computing power. I would like to add something more. We can see that computing today is not sufficient and economical, difficult to get, which leads to the fact that our businesses, governments, universities, and research institutes do not have sufficient computing power. When we prepared this slide, my colleagues provided me with a lot of examples. I said, why not use Huawei's own case? So this is a case about Huawei. The number of AI training operations is over four thousand on a daily basis, which means over 32,000 hours daily. But still, there are a lot more on the waiting list in our ModelArts system. There's another anecdote to share with you to tell you how our doctors work well. They would not do the training during office hours. So before they leave the office, there was a lock on the ModelArts and put their model on the model. So we would do the model training at night so that our other colleagues can do the training during daytime. The next day, we will see the results of the training after a whole night of calculation and training. So even if we have invested a lot in model training, the resources of computing power are still scarce. We use Atlas 900, provided in the form of cloud AI cluster service. This is the fastest AI training cluster globally. It is based on Ascend 910 NPU. Through Huawei's integrated communication, bank, and resources scheduling system, we utilize ultra-high-speed training data cache to fully unleash the strong power of Ascend 910. Now our computing power has reached 250 P, which is equivalent to 500,000 pieces of computing power. In the future, we can expand this cluster system on a linear basis. Mr. Hu also shared with you another figure: that is, within 59.8 seconds, there is a typical network of ResNet-50 on top of ImageNet-1K dataset. We have the fastest speech, which is over 12 percent faster than our competitor. Our cluster can unleash such a big mark on computing power. It's because our engineers have optimized the system from software to hardware, especially the gradient synchronization and parallel to many batch computing. From TCP switching to Spine switching, we have achieved all connectivity and obstruction-free high-speed switching network. So there is virtually no delay in our network. We also highly collaborate our chip, software, and hardware, which would unleash abundant computing power. Number three: AI service should be continuously evolving. Traditional IT is based on clear demand and rules, featuring high certainty and the separable development and production. It needs a typical supplier and demanding sites. But this should be changed in the AI era. In particular, deep learning is based on statistical computing. It is based on the model of infinite data set development, which means it may not adapt to all the changes in the environment. Therefore, the key is to build a closed-loop online system including production, operation, development, and training. With such a closed-loop online system, we can utilize the online learning capability to enable the model to adapt to the changing environment continuously and optimize continuously. In this sense, it will become a truly evolving AI service. We often tell our clients, well, if we were to ask to build a system on-premise, it is not the optimal choice because we need a continuously evolving AI, and your system is based on infinite data set development. We created and developed ModelArts, which provides a continuous interaction framework. We provide device and collaboration capability to accelerate the AI evolution speed of enterprises. For the automatic data reinforcement function, for example, when we train the OCR document model, a few pictures can be automatically expanded to over 1,000 pictures to achieve the same progress speed of training. We can save up to 80% manpower. Also, based on our model self-optimization and stop-tuning framework, and based on our multi-dimensional search engine, the result is 8 percent higher than the result of expert model and 6 percent higher than the best performance of automatic optimization. From 91% to 97% in auto mode. In the autonomous driving sector, the ironclad capability can enable data updating of models compared to previous once-in-a-week intervals. On the third day of Huawei Connect, we will announce our new ModelArts 2.0 version with new features, and I highly recommend you to pay close attention to it. And now I will talk about the fourth key factor: organization and talents. The alignment is an active process. We believe that AI should think digital and act human. The wisdom of AI is the digital manifestation of human wisdom. For AI developers, we provide them with a threshold that is easier to use and close to the industry. The threshold should be lower. In combining the AI and industry, we provide ModelArts, a one-stop platform for AI application development. It helps users to build and deploy models and manage the flow covering the entire circle. And for the experts in AI, we are thinking about the collaboration of AI and experts in implementing AI. We provide transparent decision-making rules so that AI can play a better role in assisting decision-making, so as to make the suggestion more acceptable and efficiency higher. It starts from the changes of enterprise organization and process. We cannot use AI by copying traditional IT models. AI needs matching talents, organization, and process to fully unleash efficiency and effectiveness. Now please watch another video. It's about the feelings and observation of experts from Shenzhen Airport.

We have shared with you four key factors to implement AI, which are clearly defined business scenario, readily available computing power, continuously evolving AI service, and organization and talents. Today, we officially announce Huawei Cloud Industrial Intelligent Twins. Huawei Cloud Industrial Intelligent Twins is the industry-oriented smart solution by Huawei. It is an industrial smart upgrading engine. To be more specific, based on the operation strategy of industrial operation, it integrates technology of device, edge, and cloud. Through intelligent recognition engine, intelligent prediction engine, and intelligent optimization engine, we can intelligently analyze and process information and drive smarter production in the physical world. In the industry sector, AI application needs the support from our partners in all walks of life. We hope that our partners can work together with us to accelerate the implementation of AI projects utilizing Industrial Intelligent Twins so that our vision can become a reality. For example, in our cooperation with PetroChina, we use intelligent recognition engine to support reservoir identification. The result is that the reservoir identification time has been reduced by 70%. We worked with Salient Hope to use intelligent prediction engine to dynamically match customer demands. The customer matching efficiency has increased by 28.5 percent. We also worked with Golden Stone Group to use intelligent optimization engine. We introduced AI into the blending and cooking coal quality prediction process. In this way, AI can become any silver bullets for the blender, and we managed to save 10 million per million tons of coking coal. Huawei Cloud Industrial Intelligent Twins can be implemented in different industries, including energy, mining, coking, power, cement, and synthetic fiber, which surely introduce AI into industrial scenarios and accelerate the intelligent upgrading of industry. Today, we are happy to be with us Professor Gong from PetroChina, who will share with us the case of Huawei Cloud in the manufacturing industry. We will see how far we can use Industrial Intelligent Twins to empower oil and gas exploration. Welcome, Professor Gong.

Good morning, distinguished guests. It is my delight and honor to attend this conference, and I want to thank Huawei for the kind invitation. What I'm glad to share with you today is CNPC's joint exploration with Huawei on the power of cloud, the transformation and upgrading of oil and gas exploration and development. Over the recent years, CNPC or PetroChina has aspired to boost its competitiveness and transform itself into a world-leading comprehensive global energy company. So it has launched the informatization strategy of a shared CNPC. To date, in the domain of exploration and development, 15 information systems have been developed, and the consolidated standards for data models have been defined for 450,000 wells, 500 reservoirs, 7,000 exploration areas. Historical data over 60 plus years have been put under centralized management. And the development and deployment of these systems have significantly advanced the company's efforts to reduce cost, raise productivity, increase reserves, expand capacity, enhance efficiency, and remodel its production information sharing and digital transformation. Moved up forward, let's look at a certain data. In 2018, China imported altogether 440 million tons of oil, raising China's reliance on imported oil to 70%. PetroChina is tasked to ensure China's energy security. Following national strategies, Chinese oil companies have stepped up their exploration and development. But we also know that there is a lot of pressure on such efforts. In recent years, first, it is increasingly difficult in plain terms. The easily recoverable ones have been already done. According to our data, over 70% of the proven reserves are of mid-low, if not ultra-low, permeation. That means there are very high-quality ones. The water content of the developed oil fields is approaching 90%. Therefore, both the reserve and production face great pressure. Second, the profitability is hard to control. The cost of production remains high. Third, environmental protection and green development have posted new challenges to our production. Therefore, so far, we are facing a series of difficulties. A lot of the Chinese and international experts have carried out research and come to the view that AI represents an important solution to many of the challenges in oil exploration and development as the best possible technology and methodology. Our global peers have joined hands with IT giants in recent years. For example, Shell with Microsoft, Total with Google, they have expanded their application of AI and other next-generation IT in exploration and development so as to digitize their business. PetroChina is visionary enough to partner with Huawei and introduce Huawei's Industrial Intelligent Twins into our technology. So with knowledge graph and natural language processing, machine learning, and other AI technologies, a new system is built and computing applications are made. So we are providing intelligent analytics to learn the production management so that we can also expand our reserve and production. The decision-makers can identify patterns from mega data and improve the efficiency and sophistication of operation and management. I trust that this will be a great endeavor to jointly bridge the divide between applications and business. Hereby, I would like to present you the cognitive computing platform we have co-developed for exploration and development together with Huawei. This is a shared and scalable AI compute platform designed around four key factors: only data, algorithm, computing power, and the scenarios. Data processing, machine learning model launches, and the inference applications are all hosted in this one-stop AI development environment. The development and deployment of this platform provides business innovation with intelligent drivers and an ecosystem. It enables knowledge to be preserved, inherited, and shared. Through a cognitive computing platform, a knowledge graph covering all functions of development and exploration is taking shape. This is a self-improving and organically growing project. The seed of knowledge is burgeoning and gathering the power of more knowledge over time. 10, 20, 30 years later, it will definitely grow into a towering tree nourished by intelligence and further into a forest of knowledge, revolutionizing oil exploration and development with technology. Next, I would like to give you two examples on applications which were touched upon by Mr. Zheng. I would like to give you further information. Oil exploration involves extremely complex science. A critical step is well logging, whereby in geophysics, we, in the course of drilling, use all kinds of ways and instruments and analyze resistivity, natural potential, acoustic velocity, and many other data, and identify oil and gas reservoirs. Experts matter, so we need a team of experts to take the lead in a completed job. In Daqing oil field, by a cognitive computing platform, the people there applied machine learning to 900 wells, collected data about logging, realized the intelligent recognition of oil, gas, water horizon, shortened the assessment time by 70%, and accuracy reaching the expert level. I believe that with more practices, the accuracy will be even higher as a result of learning. By using knowledge graph, professional threshold was lowered, expertise inherited, and efficiency remarkably improved. In a second case, which is about oil and gas production, the use of IoT and machine learning has allowed oil fields to receive quantitative diagnosis and remote real-time online management. So in the traditional oil fields, the machines need to be attended manually. Now, by intelligent recognition of irregularities, we have improved the accuracy of such recognition above 90%, and the oil field management operated from post-accident diagnosis to pre-accident alerts. Maintenance cost went down by 20%. Eight oil fields in Qinghai province became unattended. You know, the area is in a very barren piece of desert, and the center of gravity for management has been relocated from the desert to the Dongsheng and plains. More than 700 people in the management were relocated, and 800 field workers were reassigned. This has reduced the production cost and improved the sophistication of oilfield management. Finally, I would like to note that PetroChina has completed the planning and designing of the intelligent oil fields and explored AI in 22 selected scenarios in exploration and development. The seeds of intelligent applications are budding and will soon grow into towering trees. It is my conviction that in the near future, in partnership with Huawei, CNPC or PetroChina will constantly improve its cognitive computing platform and upgrade its oil exploration and development business. Thank you.

Thank you, Professor Gao. Professor Gao's presentation is very enlightening. We also hope that our Cloud Industrial Intelligent Twins will help more companies to accelerate their intelligent upgrading process. Huawei Cloud hopes that we can work together with a growing number of companies to cross the chasm of AI commercialization, so that AI can really benefit all sectors and industries and bring new commercial success to our partners. A growing number of our customers are turning to Huawei. In the beginning of this year, after one day's brainstorming, we were thinking about our slogan, that is, Intelligent Future. With that, I announce my presentation is concluded. Thank you.

Thank you, Mr. Zheng, for your expertise on Huawei Cloud AI for industries. At the same time, many thanks to all of our speakers for their inspiring speeches.