Huawei Connect 2018 Day 1 Keynote Live

Ladies and gentlemen, the conference will begin in three minutes. Please be seated and kindly mute your mobile devices. Thank you.

Ladies and gentlemen, the conference is about to begin. Please be seated and kindly mute your mobile devices. Thank you.

When life starts, intelligence falls. Our world expands with our desire to explore. We feel the peace of life, the happiness and passion, love and courage, and those worlds beyond imagination. We will answer an intelligent world. Cloud, edge computing, and devices are advancing day by day, working faster. Intelligence starts in the mind and is built into the hardware around us. So everything will be connected. Well, since the era of AI is here, let's work together to bring digital to every person and organization. Intelligent world, activate intelligence.

Distinguished guests from around the world, welcome to Huawei Connect 2018. Please welcome Mr. Eric Xu for his opening speech.

Listen to me, the details of our full-stack, all-scenario solution, which is a pure technology topic. Thank you for your time. First, let's take a quick review of what the full-stack, all-scenario solution looks like. As Eric just mentioned, it includes four layers: layers of chip, chip enablement, framework, and application enablement. I will go through all these layers one by one. First, let's start with the chips. At the beginning, the first question came to us is: what kind of chipset should we develop? What's important to our customers? Because different customers should have their own way to use AI differently. And for us, different applications of different organizations prefer different deployment scenarios. So we believe each scenario is unique, and all deployment scenarios by different companies or organizations are equally important. So we define our goal, and we believe we should be able to develop such a kind of chipset that can provide optimum performance but with minimal cost in all scenarios. That's our goal.

Well, it's not an easy job, as you can see from the table list here. But just to show a few application examples of AI for different scenarios. As you can see from the table, if you look at the computing capacity, it can be from 20 tera-ops to 200 tera-ops for the cloud, and 20 mega-ops for the earphone. This is 10 million times of a difference. And if you look at the power budget, which is also important, it can be as small as just 21 milliwatts. It also can be more than 200 watts for cloud applications, or even higher. And if you look at the model size, sometimes it can be as small as 10 kilobytes. Well, for cloud applications, even 300 megabytes is acceptable, or even higher. Again, it's 20,000 to 30,000 times different. There is also a more than 100 times difference of latency requirements for AI. There are two kinds: training and inference. Inference is 20 times different. It's easy to say that inference will happen everywhere in any scenario, but people may argue whether training is needed in those scenarios other than cloud and edge. Our point is yes, because we define customer privacy protection as the highest priority. Whenever there is a customer privacy concern, local learning is needed. To cope with this huge dynamic range here, the different requirements, we prepared a series of IP and chipsets, all the way from Ascend Nano, Tiny, Lite, Mini, and Max. It's a series of chip and IP design. This doesn't necessarily mean a multiple architecture design, but whether it be for a unified architecture is a critical decision to make. Because the benefits of a unified architecture is quite clear: the developers only need to develop the operators for one time, and this developers can have enjoyable, consistent development experience. And gained experience in the development for one scenario can be migrated to another. Now the service runs seamlessly. It's fantastic. But the challenges because those unified architectures are also unprecedented. First, you need to have a high computing scalability. You may have two options. One is a top-down approach, that means a design or architecture optimized for the smallest computing scenario and rely on stacking up to meet the biggest computing scenario. With that approach, we found the unacceptable power dissipation and chip area will be unavoidable. While there is a second option, a bottom-up approach, which means you may design your architecture optimized for the biggest computing scenario and miniaturize the smallest scenario with fine partitioning. If with that approach, then you need to be ready for the very complicated task scheduling in software design. And for memory, both the huge difference in memory bandwidth and also latency has to be well aligned with varying computing power adapted to different scenarios. And also, interconnections have also to face the chip area and the power dissipation constraints. What's our choice? Packaging by years of successful chip design and unparalleled customer understanding, we go for unified architecture. The unified Da Vinci architecture, as we named it, was so scalable compute, scalable memory, and scalable interconnection are the top three sets of technologies that we developed to make this architecture unification possible.

To have a high scalable computing capability, we first designed a scalable cube as our high-speed matrix calculation unit. With its maximum configuration, 16 by 16 by 16, one cube can complete 4,096 matrix operations within one clock cycle. Within one clock cycle, that's it. Okay, with more important standard at 16-bit and 16 by 16 together, we say it's computing capability and efficient data stacking capability. Now we are ready to use a single architecture, single unified architecture to support all scenarios. For example, for the computing scenarios, let's say a smartphone may need less computing power, and generally also need less power budget. The cube can be scaled in to 16 by 16, down to 1 by 1. So much less computing power, also much less power budget or power dissipation. And together with the one instruction set, this flexibility successfully provides the balance of power, computing power, and power dissipation. With multi-precision also supported, the most efficient calculation for each different task becomes possible. Just what becomes possible. And given the extremely high density, high computation density, the integrity of power supply becomes also critical. When the circuits are running at full speed, the specialist tanks or picosecond power current control technology, this extremely critical requirement is met effectively. At the same time, the Da Vinci core has also integrated a lot of ultra-high bandwidth vector units and scalar processing units. Inside this multiple compute design is not only good for calculations other than matrix, but also makes this architecture ready for potential change in the future of the types of calculation for new neural networks. And to have a scalable memory, each intervention core is equipped with dedicated SRAMs with their function fixed while capacity changeable to adapt to different varied computing power. And all this memory is designed explicitly to low-level software, thus enabling the fine control of data reuse. Okay, for further cooperation with the Ascend 910 family, which is the perfect balance of computing power and power dissipation, especially for data center application, the on-chip ultra-bandwidth mesh network connects multiple Da Vinci cores together with guaranteed low latency, extremely low latency communication between cores, in the core, with other IPs, and with the help of this 4-terabyte L2 buffer and 1.2 terabytes per second HBM, the performance of this extremely high computation density core is maximized and fully utilized. There's one thing I want to tell you about the chipset Ascend 910. Thanks to over two-and-a-half-D packaging technology, the Ascend 910 chipset has integrated a die inside, including a die for HBM compute and I/O. Thanks to this innovative architecture, as you have seen from our previous presentation, we now are ready to provide this outstanding energy efficiency across all scenarios, whether extremely little power budget or extremely high computing power is required. With this Ascend series of chips, Huawei becomes the unique in the world of first all-scenario AI IP and chips for artificial intelligence.

And the chipset is a layer of CANN, a layer of operators for the chipset. In this layer, the most important challenge or critical challenge is you always need to trade off between the performance of the operators you develop and the development efficiency, even the fact of faster deployment of AI. We prefer to define an operator as to bring both high efficiency and high performance. First, let's take a look at what's ahead. What Harry and Eric has said, the published machine learning papers has been growing at the speed of Moore's law since 2009. Eight years for nowadays, I can share you a story. You know the idea of any conference of NeurIPS, Neural Information Processing Systems, will be held in Montreal in December this year. You wouldn't believe, all 8,000 seats were registered within just 10 minutes. Obviously, this lively, active scene will continue in academia. And from an enterprise perspective, I'd like to share with you the Gartner's idea. Gartner defined a digital disruption scale measured by levels. My five levels: enhance as the first level, digitize as the second level, and transform, reinvent, and then revolutionize as the highest level, level five. And accordingly, Gartner suggests in year 2018, the main impact of artificial intelligence to enterprises is about enhance your current business, enhance your established business. While in five years' time, the impact of AI will be around transform and reinvent. So obviously, probably there has never been such a technology that can bring such a huge impact in such a short time, in such a fast way. Put all them together, we suggest we are entering into an era of dual prosperity, both academia and industry. What technically means for the layer of CANN, what we are talking about, it means that we need to be ready for the booming computing operators diversity. The diversity comes from many aspects. It could be because of different applications, different models, different networks, and different accuracy requirements, and different resource budgets, etc. And maybe many more aspects that are not known yet. Well, if you look at the development language that we are using today to develop those operators, they are either good at performance or good at development efficiency. Obviously, it will be good if we can have a new tool that can bring us both high performance and high development efficiency. And that's the right motivation of CANN, Compute Architecture for Neural Networks. And here, yes, tuned here is a high-level architecture of CANN. The key component of CANN is the highly automated operator development tool, so-called Tensor Engine. With Tensor Engine, the computer operators development, generation, optimization, and tuning can be all automated. All automated. And with the one main system interface, one DSL interface, this Tensor Engine is designed as a beginning to be used both by professional developers and non-professional developers with the same tool. We suggest professional developers to focus on those operators and extreme performance, while non-professional developers may use the same to develop whatever you want. And CANN also supports using TVM to develop our operators for Ascend series chipsets. And the other operators development can be found in the CCE LIB. CCE means Cube Compute Engine. And all non-professional developers, including you using TVM, developed operators can be found in CCE LIB. Let's take an example to see how powerful is this. Example is from Huawei internal operators development project. ReduceSum is a popular operator in TensorFlow. With CANN, it's 63 lines of code, in which Tensor Engine suggests 22. So almost three times greater development efficiency.

Then the layer of framework. As for AI framework, no doubt there is already very crowded. There are a lot of areas available in the market. But unfortunately, we did not find any of them to fulfill our requirements. And we believe an AI framework should be designed timely, efficient, that means dramatically reduce the training time and cost, and also need to be long-term efficient, such as to use the least amount of resources while delivering the highest energy efficiency. And more importantly, this framework also should be adaptive with different scenarios. It could be cloud, to the edge, could be device, could be anywhere. Okay, so first, let's see why such kind of framework should be needed. Let's see what's ahead in the game. We believe the future of AI will be highly dynamic. Not only because we noticed the trends mentioned by Michael Jordan, say the AI being supported, mission-critical, personalized, and cross-organization, etc. And that one is pushing the frontier of academia research to say, for example, AI in dynamic environment, insecure AI, even area-specific architecture. But also the continual development in industry. You know, in just six years, according to OpenAI, computing power required of a single neural network increased 300,000 times. And there are no signals to see any possible slowing down. More importantly, GDPR has applied to organizations across the world since May 25 this year. What that means? Maybe you know, GDPR is a long time ago. Our human beings, our society has mastered the intelligence to regulate a lot of assets, to regulate land, to regulate river, to regulate spectrum resource, and to regulate all the assets we have today. But for data, not yet. GDPR is the first official, comprehensive effort to regulate the data. Obviously, it's not as most at this impact is historical. So what technically it means for the AI framework? It means training or inference, we need to be happen anywhere, everywhere. So this is how we understand it. So we suggest an AI framework should be a unified training and inference framework that can enable training and inference anywhere, while keeping the development experience consistent, be it a device, edge, or cloud, be it independently deployed or cooperatively deployed, physical cooperation between diverse networks or device and cloud, or engine cloud, whatever. The MindSpore is such a kind of all-scenario native AI framework. And we are developing the complete MindSpore. It will be available early next year. And shown here is just the high-level architecture. It includes a model library, graph compute tuning, application program interface, and a device-to-cloud cooperative distributed architecture for machine learning, deep learning, and reinforcement learning. More importantly, this is all-scenario native. At the very beginning, we designed MindSpore to be able to be either very small or very big, that can adapt to different scenarios. Well, here is a smaller one, the small one, the on-device learning framework, the on-device version of MindSpore. Look at those numbers. The total size of the framework is just less than 2 megabytes. And the required RAM is no more than 50 meg. The required RAM has been reduced five times compared with the nearest solution available in the market. And this small, on-device MindSpore will be available early next year. And here is the bigger one. You haven't seen slides before. By connecting 1,024 Ascend 910s together to a unified computing cluster, the Ascend Cluster can give you 256 petaflops computing power. With that, we suggest or we believe you are able to train your model much faster than ever before. And together, we say it's 32 terabytes HBM. If you prefer, now you are ready to create a new model, maybe bigger than ever before, but much easier. Okay, the critical...

Thank you, Mr. Eric Xu. And now let's invite Dr. Vishal Sikka to share his speech. Please welcome Dr. Vishal Sikka.

Good morning, thank you so much, and welcome to Huawei Connect. The buzz here is all about AI. It is amazing to see this, to see the future that we are looking at with intelligent systems. So I want to take some time this morning to build on Eric's and Wang Tao's talk and talk to you about the great waves that are ahead of us, the waves of intelligent systems, waves of artificial intelligence, and how do we ride these waves.

How do we not only not let these waves become disruptive to us, but indeed have the fun, have the achievement of riding the great wave ahead? Eric briefly talked about this. AI has been around for a long time. In fact, humanity's fascination with intelligent systems goes back centuries. But the field of artificial intelligence started in 1956, led by these two gentlemen, Marvin Minsky and John McCarthy, at the famous conference in Dartmouth.

In the early days, in the first couple of decades of AI, a lot of breakthrough work happened. You see some examples here. The gentleman in the middle is Arthur Samuel, who wrote the first machine learning program that was using machine learning to play checkers, to play Go. A couple of years ago, when AlphaGo defeated the world champion of Go, there was big news. But indeed, Arthur Samuel had built a machine learning program to play Go in the late 1950s and early 1960s.

Right next to that, you see the first mobile robot called Shakey, that was built at Stanford and SRI. And below that, you see a picture of Marvin Minsky's own PhD thesis. This was one of the first neural networks, a physical neural network built before transistor technology, called SNARC, which was used to solve maze puzzles. And of course, on the furthest right is Frank Rosenblatt's original perceptrons implementation, also in the 1950s.

And then these two papers are papers that I wrote. The first one is a paper that I wrote when I was an intern at Intel's Artificial Intelligence lab in the early 90s. And the last one is my own PhD thesis, a one page from that. My PhD thesis in AI at Stanford was in integrating specialized procedures like neural networks and other specialist reasoners into a logic-based proof system.

So in the early days of AI, there was a lot of focus on broad application of AI and the representational aspects of AI, to be able to represent concepts using AI. And Eric referred to this, there were a couple of AI winters that followed, where a lack of success in AI led to some disillusionment, some disappointment.

But in the last five years, we have seen tremendous interest driven by some breakthrough achievements. And there are many examples of these achievements here. You see facial recognition done at very wide scale. This is a picture of a level 4 autonomous vehicle that Baidu just launched. I took a picture of today's SQuAD 2.0 leaderboard, and natural language processing systems are routinely now able to come close to human performance even in SQuAD 2.0.

SQuAD 2.0 is a natural language benchmark. The original SQuAD benchmark had human performance, there was a simpler benchmark, but the human performance in that was already routinely exceeded by autonomous systems, natural language processing systems. So now the SQuAD 2.0 benchmark is a more sophisticated reading comprehension benchmark, and already again we see systems, and there is a paper from Microsoft Research, and Dr. Hau will talk right after me, about how natural language processing systems are now able to get close to human performance.

So there is a tremendous interest, a tremendous set of achievements that have happened in the last five years. And these are a result of three factors. And the first one is advances in some techniques, primarily based on neural networks. And I showed you an example of Rosenblatt's perceptrons that was built in the 1950s. The deep neural networks, which are massive in size, augmented by techniques like convolutional neural networks for vision and image processing, and recurrent neural networks based on that are used for speech and language processing. These have led to some breakthrough results.

But really, the results are also as much a factor of the massive improvement in computing capacity that has happened over the last few decades. The computers today, and Eric talked about this, the Ascend chip with 256 teraflops of operation, this is unthinkable probably when Rosenblatt built his perceptron. The entire world's computing capacity put together was not 256 teraflops.

And so this combination of the advances in techniques, the advances in computing, and the availability of large amounts of data has led to a lot of these breakthrough achievements in artificial intelligence in the last, I would say, five or six years. And that has led to a huge set of opportunities. Today, there are several thousand startup companies funded by billions of dollars in venture capital that are addressing different parts of the overall AI spectrum.

And within the enterprise as well, when we think about this, we realize that every single industry can go through a significant transformation led by AI. In every industry, customer engagement, our ability to understand what customers are doing, to engage with what customers are doing, can be dramatically transformed using AI. AI can be the eyes, the ears, the voice of an enterprise to the customer.

And in the back-office, significant operational efficiency, operation simplification, and automation can be achieved by using AI. From simplifying many of the core processes which are labor-intensive, to in asset-heavy industries, simplifying the tasks of maintenance, asset maintenance, predictive maintenance, being able to bring a dramatic new efficiency to that.

If you think about the oil and gas industry, for example, the task of upstream exploration, efficient and less disruptive exploration can be dramatically enhanced using AI, where you're analyzing massive amounts of data coming from seismic data generation. And today, there are thousands of physicists and mathematicians who analyze data coming from these seismic events to understand where the oil or where the gas is. This process can be dramatically amplified and simplified using AI.

And similarly, the act of exploration, once you have identified where the resources are, then the act of exploring and digging and getting to that, the execution of that can be simplified with AI. And of course, the task of maintaining complex machinery in remote areas can be dramatically improved by using AI techniques. We all know about consumer applications, but in every single industry, these kinds of applications exist.

And yet, despite all this interest and despite all this excitement and opportunity, there are still significant and structural limitations that we see in AI technology. When Sean talked about this, it is still early days, and despite decades of work, we are still in the early stages of AI. And there are significant limitations that are still in front of us.

We see lots of examples, some famous examples that have happened recently, of autonomous driving not quite doing its job, famous examples of misidentification and mislabeling and so forth. There in the middle, there is a very interesting paper that was just published. It's called 'The Elephant in the Room,' where the authors actually put a picture of an elephant inside a room, and then dropping a picture of an elephant changes the identification of all the objects that are in this picture.

And of course, as humans, we can immediately recognize that there is an elephant in this room, and it is impossible for there to be an elephant in this room. And yet, of course, these neural systems don't have a way to model the world in this way, to understand the semantics of the world in this way. I mentioned earlier that in the early days, AI used to be a lot more about representation and about inference. Even though we have seen dramatic progress in the last few years, we have actually not quite cracked many of these fundamental problems around being able to understand, articulate, and reason about the world. And these are essential qualities for enterprise AI.

So when we look at the emerging stack for AI in the enterprise, we see three distinct layers. We see the hardware, where Moore's Law, which has been the guiding force behind the development of hardware over the last 52 or 53 years, has now more or less come to a stop. I was reading this interview with David Patterson, who won the Turing Award last year, where he talked about the fact that if Moore's Law was still continuing, we would be a factor of 15 ahead of where we are. So in the last few years, Moore's Law has significantly slowed down.

And the slowdown of Moore's Law has led to the opportunity to build new kinds of hardware, domain-specific hardware, AI-specific hardware, to continue the dramatic advances. And then around that, the availability of systems and cloud is creating an opportunity to build AI-specific hardware systems. And the layer above this is the layer of software, where you have two distinct categories: the engines like TensorFlow or Caffe or MXNet and so forth, PyTorch, and the layer above that of the developer experience, of the platform and the tools and the software ecosystems.

And then finally, the top is the layer of AI applications and AI services. And when we look at the opportunities and the gaps in this stack, we see that while there is tremendous opportunity, and the simple way to look at the opportunities, we have the opportunity to transform enterprises in every industry. Why? A really great enterprise AI platform that would deliver a seamless developer experience to build multifaceted enterprise-class applications on an underlying hardware platform that would deliver a full-stack experience that was integrated and yet open and high-performing, that continues to be the great opportunity.

But when I think about the current reality against this stack, we still see that there are significant gaps. When it comes to services and application building, there is a lot of, I mentioned earlier, thousands of startups, and many of them are building applications. But these applications are still scratching the surface of what is possible. The applications are not connected, they are point solutions, they are not interconnected, they are not interoperable, they are not a part of a shared common platform.

There is still a massive shortage of AI talent. As we speak today, there are maybe 300,000 trained machine learning engineers in the world, and this number needs to be in the tens of millions in the time ahead. The availability of services is still quite weak. Long-term, mature AI enterprise concepts like lifecycle management are still not there in the AI stack.

And when we think about the developer experience, the developer experience today is quite broken. There are many frameworks and engines, I mentioned some of them, but our ability to build an application today is still very fragmented. I have tried to do this myself. I have observed developers building applications, going from identifying the problem, finding the sponsors, getting the budget, getting the governance right, to getting the data, making sure the data is labeled, making sure there is governance around the data, making sure that the data is secure, its lineage is well established, cleansing it and so forth.

Getting the hardware, hardware is a crucial issue, and Eric alluded to this in his speech. There are huge differences in price-performance when we make hardware choices. On the one hand, we have elasticity that we need in order to scale our delivery of AI systems up and down, and on the other hand, we have the cost-performance, which can be dramatically different depending on what kind of a hardware choice we make. And then that is sometimes at odds with openness, because making a cloud choice gets us locked in to that cloud choice.

So how do we navigate when it comes to hardware, this choice between elasticity on the one hand, the price-performance on the other hand, and openness on the third hand? And then once we have done that, is the actual task of building the application, identifying the software tools, finding the experts, trying out the tools, getting around the fact that many of these tools are still opaque, they are not transparent, they are not explainable. We have to understand, are we doing the right thing? Are we going to, under the right circumstances, be able to use the right tools?

And then finally, lifecycle management. Once we have built and deployed the application, how will it survive? How will it evolve as the data evolves, as our business process evolves, as the data sources evolve, as people leave? How do we transfer the knowledge of what is inside this system? So we need an AI platform that has the ability to deliver a seamless experience across these kinds of steps, and today we don't have that.

And of course, finally, the hardware, I already talked about it. Again, there we have this trade-off between the elasticity on the one hand, the price-performance on the other hand. So what we really need for the times ahead to deliver the full potential of AI is a partner that can help us take, that can help us deliver this kind of a full stack across the six layers and deliver both on the one hand the economics, but on the other hand the performance and the flexibility. Can we go to the next slide, please?

So we need a new approach. We need a new partner who can help us understand and ride these great waves. Can we go to the next slide, please? A partner who can help us understand, who understands the business and the enterprise complexity. A partner who on the one hand understands the AI and what it can mean for us in the enterprise and can deal with the rapid evolution of the field that is going to happen in the near future. Who understands the limitations of the technology as well as the opportunities and can translate that into solutions.

Who can deliver these solutions across the stack, a full-stack solution, and the cost-performance of the full-stack solution to help us deliver an enlightened enterprise. I was really happy to see, I was looking forward to seeing what Huawei has to announce today, and I was very happy to see the announcement of, on the one hand, the Ascend microprocessor, that is the first AI-specific chip that Huawei has released, and the incredible performance and the AI-native CANN programming model on top of this to build optimizations.

You know, Python code, a lot of AI applications are written in Python. We can actually deliver a thousand times improvement in Python execution by optimizing it into the AI hardware. And then, of course, the MindSpore libraries for building AI applications, and then the ModelArts programming model and the developer framework for a great developer that is integrated. I believe that Huawei has done a wonderful job in creating such an open stack as well as a full-stack experience across the board.

People often ask me, what is the future of enterprises with AI? What happens to jobs? What happens to the skills issues and so forth? And my own sense is that if we look back over the last several decades, as well as even before that, going up to the Industrial Revolution, about how automation has helped us transform ourselves using these tools, I think an enterprise can achieve the same kind of benefit by the use of automation, by the use of intelligence, by the use of AI.

As systems become more intelligent, enterprises can become more enlightened. AI can help us focus and find our own wisdoms much more effectively by taking away the things that can be well-defined, that can be well-prescribed, and automating those. We are free to pursue our own unique wisdom, to pursue our humanity. And the same thing applies to enterprises. In enterprises, intelligent systems go hand in hand with the enterprise's ability to become intelligent, to become enlightened, and go a step beyond being intelligent into delivering our own creativity, our own innovations, our own ability to invent the future.

I think AI, like any great technology, like any powerful technology, can be a disruptive force, or it can be a great force for helping us improve our future. It can be a force for us to help us, a wonderful way to build our products and solutions to serve our markets, to serve our customers. And it is that purposeful AI that I am interested in pursuing. I believe that in the times ahead, AI is going to offer a great set of new waves. And just like in surfing, when we see a great wave, we can either get wiped out by it, or we can learn to ride it.

In fact, 'hang ten' refers to the act of hanging our ten toes over the surfboard, meaning learning to ride a wave so smoothly, so efficiently, so effectively that we have the freedom to actually put our toes on top of the board and hang ten. And I hope that at this Huawei Connect, you are able to really learn about all the different AI solutions and technologies that can help us to ride and hang ten on the great waves ahead. Thank you very much.

Thank you, Dr. Vishal Sikka. Now, please welcome Dr. Hau to share his speech. Please welcome Dr. Hau.

Good morning. It's my great honor to take the next 20 minutes to share with you Microsoft's vision and strategy of AI. You may all know that the reason we are discussing AI here today is from digitalization to PC to IoT, everything is digital. With digitalization, everything can generate data, and data increases significantly. Now we have big data.

Today, in 2020, your behavior every day, the things you do every day, and your experience, everything will be translated into data. With this massive amount of data and with AI and the powerful algorithms, we can do a lot of exciting and powerful applications. Microsoft is similar to Huawei, we are a platform company. We have an exciting vision because we have cloud and device, we have intelligent cloud and the intelligent edge. The interactions happen in the edge, but a lot of intelligence is achieved through the cloud. We say that we need to train the AI models, and that needs to happen in the cloud.

So what is Microsoft's vision? In science fiction, you may see the conflicts between AI and human beings, but we think AI is going to amplify human ingenuity. We can also infuse AI and make AI available to everyone, so mankind can spend our time more wisely and enjoy the interactions with human beings. So we say that AI is to help people and to amplify human ingenuity.

Take myself as an example. I've moved apartments two times in one year. I could not remember my fiction, my phone number. It's not because I don't have a good memory, it's because I don't need to do it, because the tool can do it for me. With big data and AI, we can have translation, voice recognition functions, and it can also enable us to have intelligent interaction. And more importantly, after we collect data, we can make decisions, and that is about reasoning.

Microsoft is a platform company. Our vision is to use our platform to help our partners and enterprises to embed AI into their products and services. Microsoft is just a platform in the ecosystem, but in this way, we can make AI available to everyone. And this is the way that Microsoft can make the greatest contribution to the world.

Talking about platform, recently a lot of people are talking about AI. I think in China there is the ABC concept, and I like it very much. Before AI is possible, you need to have big computing. Currently, a lot of computing tasks are done on the cloud, and some are done on the edge. AI needs a large GPU, and deep learning also requires large computing power. So this is the C in ABC. Currently, AI is based upon big data, and B stands for big data. Before AI is possible, you must have B and C, big data and computing power. And that is what we can offer through the platform.

Huawei's speakers also mentioned that chipset, hardware, and service, and then OS and some middleware, these belong to the computing platform. Above that, it is the data platform. It can process structured data. We have structured data, database, and also now unstructured data, Hadoop and Spark. They provide data lake. And then how to interact structured and unstructured data? Then we have flow data and streaming data from the edge. With data, you need to store them, and also you need to process them, and then you can conduct data mining. In addition to that, you may do statistics and make them visible. You need to conduct big data analytics before you have AI applications.

Above that, you can have machine learning, deep learning, AI framework. Above that, you will conduct experiments and debug, and then you may release it as a product or service. For any company to utilize AI to go digital, they need such a platform. Microsoft is very similar to Huawei. We hope that in our ecosystem, we can provide a platform to empower every company to easily embed AI into their products and services.

I'm responsible for R&D of Microsoft in Asia-Pacific, and I'm also responsible for Microsoft Research Asia. And this year is the 20th anniversary of our Research Institute, and this is also the 40th anniversary of Chinese reforming and opening up. So it's a very exciting time point. In terms of fundamental research, Microsoft is growing together with China to the future. We're also working on the technology transfer. We've developed a lot of teams. Early November, we will have a celebration of our anniversary. There are a lot of activities planned, please stay tuned.

There are important technologies for the future we focus on, such as natural user interface, artificial intelligence, intelligent media, big data, high computing, intelligent cloud, intelligent edge. And finally, what I would like to say is that for the entire IT industry, the rapid development relies on fundamental research, among which AI is a very important case in point. And here at this location, I would like to emphasize that AI has at least a history of 62 years. Microsoft participated in AI research for about 30 years with the initiation of Microsoft Research Institute. And we understand the importance of fundamental research. Microsoft and Huawei hope to encourage more enterprises to do research in the fundamental level of AI, so that it can lay a solid foundation for a bright future of AI application and for the entire people's well-being.

Recently, we have made some AI breakthroughs, and I would like to share with you about those. There are some important technology domains such as vision, speech, and natural language. In computer vision, ResNet is a frequently used indicator where we see breakthroughs. For the first time, our image, the computer vision capability, has come beyond that of mankind. ResNet is the deep learning module that everyone uses in computer vision. AlphaGo, which we are all very familiar with, also utilized ResNet to beat the human champion.

Then speech. In our Asia-Pacific Research Center, we worked with our American colleagues on a very important task called Switchboard, and for the first time, we have reached human parity. Machine translation, a very hot topic. In October this year, we also matched human performance in translating news from Chinese to English. And there's a very important and also challenging domain called natural language comprehension. An example: English reading comprehension means that after you read a segment of articles, you need to answer the questions with the understanding of the text. I think we all have this kind of experiences when we take English tests, and it is difficult for us to have a high level of reading comprehension.

In the beginning of this year, for the first time, our exam result also reached human parity with a very high level of exact match. Before October 1st, in the F1, the impartial comprehension part, we also exceeded human parity. I believe that with our focus and our attention on AI domains, there will be more breakthroughs in the future.

AI can help us with understanding language better, vision and clearer speech. We can also use data from different domains to deliver digital transformation. Normally, we see the top right corner part, namely transforming your product. In the past, we called it 'Internet Plus,' today maybe called 'AI Plus,' but I think it's just ABC Plus. Like if you have big data, computation, and AI technologies, you're able to transform your products and services. I think this is a commonly discussed topic.

The other three areas are somewhat less discussed, but this is where we can show our core competence. First, capability to engage with your customers. If the customer has some issues or difficulties, you need to help them to resolve those, and we call it intelligent agent. And more importantly, how can we put these potential customers as a part of our marketing team? Whether it's a B2B or B2C company, it needs this kind of capability so as to have better customer relationship.

Let's look at the lower right part. Whether you focus on sales or R&D, it is very important for you to have more efficient operations and cost. And a lot of the components are discussing how we can realize that through the ABC technologies. And last but not least, empowering your employees. I think at the very end, the value of the company is its employees. Only people can come up with new ideas, new strategies. How can we use these A, B, and C to empower our employees so that they can do a good job, not only at work but also at home? Sometimes you have some family emergency, you have to rush back, and when that happens, how can you use ABC technologies to make the employees more efficient to deal with such situations? And more importantly, here we have computing, big data, and AI. How can we motivate employees' ingenuity so that they can come up with new ideas? In this way, the company will be more competent. I think this is a very important direction.

Microsoft in China works with its partners, including Huawei. How our customers and partners can be offered better services? Today, we are exploring the digital transformation as a service. I understand that you come from different backgrounds, some technologies may be understood differently among yourself, but I think first of all, the consulting service is very important. And that's why we established our Microsoft Innovation Partnership, so that we can communicate between companies. In particular, Microsoft can discuss what the other company is about, our products, but most importantly technologies, and also future technologies. Like in one, three, or five years of time, what will the future technologies be? With that understanding and that insight, the companies will do a better job in digital transformation, and they can have a multiple-year planning.

Apart from building a platform, Microsoft welcomes our partners and other companies to utilize our platform for digital transformation. However, some other companies also hope Microsoft can help them with the last mile, the industry innovation. Because of that, we work with our strategic partners in that domain. Let me give you some examples. First, intelligent logistics with Ocean Network Express. In the past summer, I worked with them, and with the reinforcement learning, which is cooperative, we find ways to better arrange the containers that are vacant, because it's very important for us to reduce the empty container time. During our first phase cooperation, we utilized this cooperative reinforcement learning technology to reduce 10 million US dollars in terms of cost. And with the second phase up-and-coming, there will be more cost reduced.

Another example is about education in China. Learning English, it's most efficient when we are really learning from a tutor. However, if we're not able to have this face-to-face learning experience, then we can't think of something else. Here we have developed a robot that can communicate with you, that can keep scoring you, giving you feedback about your pronunciation, intonation, word choices. And at the same time, the agent can help you note down the difficulties you have in learning English. And the next time when you go to a classroom, the teacher would know instantly what kind of issues or difficulties you have. You may have heard about Longman Dictionary. We now work with Longman, and in the very near future, for every Longman textbook, there will be a corresponding English learning robot, so that we are able to be more efficient in learning English in the future. So this is another example of how we work with our partners.

Finally, about the corporate social responsibility. We understand that there are some anxieties about AI technologies, like would it be controlled by very few people in the future? Will I compete against people in the future? Or will it lead to data bias and other problems? Or even will it replace most of the jobs? Microsoft has been active in many domains, and we are proud to say that we welcome more people from the industry, from the government, to participate in dialogues to drafting the AI-related rules and regulations, so that AI would not only be pervasive but also be doing good things, ensuring fairness, eliminating bias, protecting our privacy, and so on and so forth. And we call it AI ethics or AI for partnership. The essence of those concepts is to let people trust AI more.

Today, I understand that the audience are the technology enthusiasts, but some people do not know about these technologies. When we talk about using an AI technology or other technology to change the world, we need to ensure that this technology is trustworthy, it's regulated. That's the only way. A lot of the high-tech companies think that the rules are unnecessary as it will impede innovation. However, Microsoft thinks that rules are very important, because if we're able to have some fair, open rules, it will put us on the same footing. Otherwise, without laws, some may start them early, and some tend to be conservative as they wait for others to have started. Instead, we hope that this kind of AI rules can put us on an equal footing, so it can start together, and eventually, it would be better for mankind, for our welfare.

Therefore, Microsoft's goal is to ensure AI empowers us all, that we can try to eliminate all the related negative impacts. We're very happy that today Huawei is holding this Huawei Connect conference, and we were honored to work with Huawei. Our platforms, respectively, Microsoft Azure and Microsoft AI, Huawei has also provided us with tremendous support. Finally, I'd like to make a little advertisement here. This is the QR code you see. If you scan it, you're able to access the Microsoft website to learn about our latest AI development. At the same time, two weeks from now, namely at the end of October, our Microsoft Tech Summit will be held here in Shenzhen, in this Expo Center, where we will launch our latest products and offerings. And we hope that you can also join our event.

Many thanks to all of our speakers for their inspiring speeches, and thank you all for joining us this morning. Over the next two days, we will continue to explore more exciting topics together. Here's a preview of tomorrow's keynote event.