 China Mobile AI Strategic Policy China Mobile is at having AI full coverage on key capabilities in products, creating perfect technology innovation system, promoting the deep convergence of AI in the real economy, accelerating national innovation construction. AI Application Pioneer Following the principle of high goal, high starting point, high standards, China Mobile proactively designs AI layout as it emerges. By introducing AI capabilities in several fields like network operation, customer service, product innovation, and management improvement, China Mobile builds leading ability in intelligent operation in intelligent service, becoming AI application pioneer in the industry. AI Industry Enabler Facing the future, China Mobile constructs the Intelligent Information Infrastructure System, which combines perception, transportation, storage, calculation and processing, enhancing intelligent application ability generally, and becoming AI industry enabler. With an international perspective and a desire for ICT convergence, China Mobile participates actively and contributes a lot in open source communities. Faced with the next generation network, China Mobile is introducing Open Network Automation Platform, ONIPE, which is the intelligent brain of SDN, NFV, to realize automatic network orchestration and management. CNCC outputs several influential use cases, such as VOL, such as VOLTE and CCVPN use cases to ONIPE, collaborates with other active members in the community and continuously contributes code to core projects. As the biggest open source community in LFN, ONIPE expands rapidly in the last few years, owning more than 100 members and over 70% global users. By taking advantage of the strong orchestration ability of ONIPE, intelligent scheduling, agile deployment, dynamic adjustment, are changing into reality. Okay, thank you. Thank you for all the support. So the topic I will cover today is practice, where's the title? The Network Intelligence and Open Source, the Practice and Thoughts. I feel like whichever conference you go, no matter it's speech, natural language, and no matter it's which industry, kind of all we all united on their artificial intelligence. So I feel lucky to be the first one to talk. I don't need to worry about going to repeat someone else's points. So we'll be two parts. I will share some practice inside Time Mobile, also how we see in the industry. And then the second part I will share some of our thoughts, I'm thinking are the things I think we can collaborate on. Okay, if we look at the standard organizations, there's tremendous work recently done. Try to put AI into the standards from 3GPP, from ITU, from ENI, and also others. So without going through the details, and we can see where the industry is going. So if we look at the LFN members, if you talk to the top officers, the developers here, everyone seems to have some part of to do with AI. And we have a couple of keynote talks on DevOps. I'm thinking as we, for now, at this moment, we have invested quite a lot on AI ops and it's next step for DevOps. I think it's, no matter it's the web skill, internet companies, so techos or vendors are invested heavily on AI, how to use the artificial intelligence technology on operation. Okay, with that, from STU, from the industry in general, I give you an idea how the AI R&D looks like in China Mobile, what is our strategy? So we don't see us as a kind of core, deep-down technology inventor. What our strategy is driven by value. Because we have been into big data for a couple of years. I mean, we all think big data will lead to big value, which is not necessarily true. So when we come to the era of AI, we think we have to derive the value, the commercial value, from the technology. So we are value-driven AI strategy. We're trying to push out the larger-scale applications to improve our business, improve our productivity. So come to details. So we have our general platform, AI platform, support R&D and production. Company-wide, and then we have the bottom, the infrastructure for the hardware, how managing the clouds, and then have the layer of common capabilities on top. What we focus on is the three major areas. It's artificial intelligence for the network, how to make our network intelligent. And the second part, the marketing, service, security, and management. I don't think I will go to details, but I will focus on the first one, the network. Just give a couple, there are tons of work to be done. So I'm thinking I will just mention what the systematic view of that work looks like. So for network intelligence, we put into three layers. If you look at the outside circle, which is the intelligence service, we can do the, for example, the business management industry services, how we improve our user satisfaction, those are on service layer. We think it's easier to do compared to inner ones. If you look at the middle ones, it's intelligent operation. I think we all put in quite lots of efforts into the lines of business and the core one is how we make our core the foundation of the network intelligent. The routing, slicing, energy, how to save energy, how we use AI to make Mac intelligent. So given that overview, and I share a couple of things, just take a few examples to give you a sense of what we did. For example, for customer care, we have this chatbot in deployment since 2014. This is a chart how we improved from last year. For now, the robot 100% to the robot to handle the customer care volume 22, almost 23 point percent. This is how we evaluate this, not just the accuracy of how we interact with our customers, how accurate it is, it's more of how many channels we put in, how many chats we have per month, and the rejection rate, our accuracy, how that impacts your customer satisfaction. That's how we evaluate the technology zone. On the bottom, it relies on speech recognition, rely on sentiment, rely on lots of NLP, natural language processing, speech technology, knowledge base, data analytics, and some part of computer vision technologies to integrate everything into. We have a closed loop circle and every single day we have data annotated, the data have our performance evaluated and confirmed by human and come back, and we upgrade the model on a daily basis. That's the first one. Give you a second example. It's a little more relevant to network. So we all get consumer and customer complaints coming in. That's a big worry of all techos. So this one, what we did is from the customer give us the call, we get the speech, we transcribe, we transcribe into text, when we analyze the text and analyze what's the problem, the concern is, we extract all the facts, then we combine those, and then that's part of the understanding what the users complain is. The second part is we tackle into all the major data we need to connect with, and we're trying to analyze what the problem is, how to solve it. So basically as we build a network-operated maintenance kind of knowledge base, and we also connect to all the data. I have layered here, there are many of them. You can imagine there are a heavy layer of engineering to connect all the data points to make the decisions smart. This has been deployed in large scale and saved us quite a lot of money. And another one is connected to wireless, it's intelligent coverage optimization. This is a back office tool used inside time mobile. A couple other things for future networker, next generation, they're not in large scale deployed yet, just to give a couple of examples. This one I presented last year with Huawei and Vodafone and the cross layer, cross carrier VPN, and this year we showed this demo, we also partially deployed it in China mobile, we combined this own app and AI on the edge. So basically the scenario is you imagine you have a security monitoring system that you have your detects. If there's a normal speech or a signal, sound signal, then if there's extreme situations from the picture, from the video, and then we notify own app and adjust the bandwidth automatically. So it's getting AI to initiate the orchestration. So this is AI for network slicing, this is one of our main thing for developing to prepare for 5G. This is AI for smart edge, without going to details, I think for this community, there are tons of work in the community and inside each of the companies. We're thinking how we can systematically put AI into own app for the data analytical part for the data analytical part to close the loop. So I gave a couple examples, but which is I don't think it's an important part of my talk. So to summarize, we have a lot more, hundreds of AI enabled systems functions. Now we have developed and deployed in the past couple of years to help improve the complex network. There are too many, so I just give you a sense. So they're in Chinese, that's intentionally I want to, because it takes lots of time to translate them. I don't think there is a key point to show them. So then what's the problem behind? We built, we spent lots of energy and human resources building this intelligence system to help a network. We see what we got, it's a great technology. I think it's a way of AI is transforming us, is transforming the industry. So we are also, at this moment last month, last week actually, we announced that we are hosting a nationwide network intelligence competition in China. So whoever are interested, you're really welcome to give a try. And I think the second part is really I want to share with audiences here some of our thoughts, what we can take, what we have as a now for network intelligence to the next level. So what we should do, what is the problem? So first one, what is a correct definition for network intelligence? So even people saying they're not, so we all talk about AI, but everyone's AI is different from others. So for me, I give you my personal definition, what is artificial application? So we can, you can have, AI means lots of technologies, you can machine learning, deep learning, transfer learning, reinforcement learning. You can have lots of learning, you can have knowledge base, you can have inference algorithms. I'm thinking in a way, so when you put everything together into an AI-based system, the system got to have a feature, it can improve itself on its own. So I'm thinking that it has to be a living thing. So if you have the system in the performance today as one, and tomorrow it's, or a week later, if it can become 1.01, it improves on its own. It has a sign of as a living body. So I'm thinking that that kind of application we call the AI. If you deploy technology, you never touch it, it doesn't improve on its own. Even it has speech recognition, even it has image processing. I use AI technology, I don't think it self is AI application. So I give you a definition from my personal view, but I'm thinking as we truly need a definition from there, we can try this too, because there are just too many cases, scenarios in our business. So we can put it into, we can use AI, AI definitely helps, but there are lots of efforts on kind of low level, high in food, how can improve that, it's, I would think if this gets clearly defined, it will be a big help. We take the intelligence to the next level. The second one, can efforts be systematic? I show you what we have done on the transport network, on the core network, on the operation on the services side. So I'm thinking where, how can make it more systematic? So we can, if you compare this concept to the levels of automatically driving cars, so you can, from the easy to hard, what is level one, what is the level five? So where, where we are now? The second way is we think it's, it's the service, the operation layer, the basic, the foundation of the network, and wireless to core network, wireless transport in the core network. And so we all know for, for the tackle industry, we need lots of planning and we construct our network, we operate, we optimize the network in different phase of this, we need AI to come in to help. What those AI technologies look like? What the framework should be? I'm thinking it's still an open question. And when we do AI, we need the steps. You need to sense what's the data and to get the data, how we can do the sensing right. So when we, we are proud, we get lots of data, but lots of data definitely means a, big data means a big cost, doesn't necessarily means a bigger value. So how we can correctly in, in a cost effective way to sense our network, the qualities for network to sense user experience and how we correct distorted data, analyze the data and then do the prediction. Those are our open questions. The third question I want, hopefully we all can work together on this, is on what level we should share data. So we cannot imagine the data process, the image processing, the CV, the computer vision. Technology can come to where it is now without image data set. So we need for AI, everyone knows it's, it's data to generate the intelligence. So where the data come from, in what way should come. So data is issue. If you look at the research papers from the network literature, you can see most of the researchers are used, you use the data from many years ago, doesn't really necessarily represent the problem of what do we face in our industry. So that's a bigger gap there. So that's why we host this competition in China. We invite developers to share some of our data. We also, you come to our, our own computing platform. You can run your things for the big side of the data, chunk of data you can come to our platform to run your computation to try things, try your own algorithms. Next level is the model. We all know now the intelligence is a narrow sense and it's not a general intelligence. It has to for a specific problem. But on the model level, how we should share? Well, it's pretty hard to share. If I have a special recognition model, I give it to you to run recognition. Probably it's easier. For network, it's we don't know which level, which problem, we can share that on model level. Another way is for now, for the last year, the big thing is in the AI field is BERT, B-E-R-T, the pre-trained model from Google. Basically it gets a level of natural language processing to find a general way which can get the meaning of the words of the sentence in a way from large amount of data. And you can take that model, they call pre-trained model, basically to catch the basic semantics of words of sentences. Then for all the downstream tasks, no matter you do sentiment analysis, you may do customer care, dialogues, whatever you do, you can use this pre-trained model and easily tune for whatever the specific application is. So I'm thinking in the telecom intelligence world, do we have this level of abstraction? Can we build this kind of model? Is it possible? If so, how we do it? So the third level is, since we're not working in intelligence, if you get the data, the data is static. So lots of problems need a environment, not much data. So what that environment should be and should we collectively build that one to foster the research in R&D and even inside the cooperation, what that environment should be open to the developers. The problem of abstraction, I showed you, you can build millions, hundreds, thousands, millions of applications, but you cannot handle the costs. You have to pay so much for building those capabilities. So I'm thinking if we look at the typical problems, if we do a level of abstraction, we get to know what's the nature of the day, of our problem. We face a classification, it's a decision. What's the foundation? What's the nature, basic nature of the problem? If we get that layer of abstraction, I think we probably can build something general to fit all the scenarios versus we're building one for each. And another way I'm thinking is, we now we borrow the algorithms which are originally designed to process natural language, to process speech, to process images, videos, but they are not fundamentally designed for network problems. Are they perfectly okay for our own problem? Or we need something special? You cannot imagine without image processing, there's no convolutional neural network. If there's no speech, there's no LSTM. So the problem itself probably can lead to another level of invention for AI. I'm running out of time. So efficient way to improve collaboration between industry and academia. I have seen this really big gap between what the industry is doing and what research professors are doing and our research. Is there open source? Is open source an easy way for this level of integration compared to the commercial software? So fifth question I want to share is, AI is not only that piece of technology, it impacts our own way of we run our business. So methodology and process to efficiently deploy AI enable function. I mentioned the closed loop, how to make your AI system really a living thing. It's not only algorithm, it's not only data, it's your business process has to close that loop to give it a life. And sixth question I want to share is, way of business and organization, how to match this new technology. Every piece of technology, especially I think it will transform the way we think and also the way we run our, we organize our folks. The last one is, I hope our network can be simpler and I hope AI will help here. Okay, that's all my content. Thank you.