Huawei's Xu Wenwei: AI is a top student in the exam room, but a poor student in the application scenario
▲Click above Leifeng.com Follow
Text | Zhao Chenxi
Report from Leiphone.com (leiphone-sz)
The three-day (January 19-January 21, 2019) 2019 EmTech China Global Emerging Technology Summit is coming to an end. According to Leifeng.com, this year's conference revolved around eight major sections: the AI era, computing power connecting the future, the finite world, the infinite blockchain and financial technology, future work, rewriting life, the digital world, and space.
The advent of the AI era has accelerated technological change and changes in the market development pattern. Anand Sanwal, co-founder and CEO of CB Insights, said at the conference that the biggest change brought about by AI is that technology is developing faster and faster and its impact is becoming more and more far-reaching. Factors that may have helped us succeed in the past no longer work. For companies, speed is very important. For example, the entire life cycle of companies in the S&P 500 index in the United States was between the 1950s and 1960s. Usually, the life of a company was more than 60 years, but now it may be less than 20 years.
Xu Wenwei, director and president of strategic marketing at Huawei Technologies Co., Ltd., also expressed his views on the current status of AI. The arrival of AI will bring about changes in three aspects. First, it will change the organizational and personnel structure of all industries. Second, boring and repetitive work can be replaced by AI, but a large number of creative work is impossible. Third, it will be used in the industrial industry, for example, to support various applications such as IoT and Internet of Vehicles, and to adopt artificial intelligence in 5G.
The maintenance cost of telecommunications networks is now 3 to 4 times the equipment cost. The network is becoming more and more complex, and 70% of network failures are caused by human factors. After adopting AI, more than 50% of potential failures can be predicted. Therefore, AI can reduce the operation and maintenance costs of 5G networks and ensure the improvement of network quality.
AI has many advantages, but there are several problems with it. Xu Wenwei believes that these problems are mainly: Problem 1: The computing power is too expensive and cannot be used. It looks beautiful but is too expensive to use. Problem 2: We don’t have good data, so we don’t have good AI. Data needs to be processed to ensure data quality. Without artificial intelligence, there is no intelligence. Problem 3: Training is too slow. Problem 4: Application scenario. AI must have a model. The debugged model is tested or tested in a certain scenario. It is a top student in the examination room, but a poor student in the application scenario.
Xu Wenwei pointed out that AI still has some flaws. First, its scope is relatively narrow. Second, it is more like research than engineering. Third, it works. If it is applied to other application scenarios, the data may not be feasible. Fourth, a good model may deteriorate over time, and there is some uncertainty.
Finally, Xu Wenwei mentioned the grim situation facing the AI field. First, there is a shortage of talent. Second, AI involves data security and privacy protection. How to protect personal data and privacy is worth further study.
The following is the transcript of the speech by Xu Wenwei, Director and President of Strategic Marketing of Huawei Technologies Co., Ltd.:
Hello everyone! Thank you very much for attending today's exchange, and thank you EmTechChin for inviting me to share with you as an enterprise or industry.
As we all know, AI is very hot. According to incomplete statistics, there were at least 300 AI-related conferences last year. Therefore, it would be impossible not to talk about AI at a conference now.
Why is it so popular? You must have seen many good application scenarios of AI. In fact, AI is not a new thing. As you know, the concept of AI was proposed more than 60 years ago. At the same time, because AI is so attractive, people have too high expectations for AI, so it has experienced two major winters.
What is winter? It means that our expectations for AI exceeded the engineering capabilities of ICT at the time, starting in 1971 when Intel invented the processor.
With AI so hot right now, will our industry experience a winter? We hope not, after all, we are just getting started. There are many people here who make investment decisions, including senior executives of companies. So, what I want to share with you today is how an investor and a senior executive of a company can grasp the rhythm of the industry and make some references for future investment decisions in the industry.
This time I share my personal views on this industry.
As we all know, there are 26 kinds of GPT general technologies in human history. What are its characteristics?
-
Feature 1: At the beginning, it was very imperfect and had a lot of room for improvement
-
Feature 2: Multi-purpose, can be used in various industries
-
Feature 3: Spillover Effect
So our earliest GPT can be traced back to the wheel invented more than 1,000 years ago. More recently, as we all know, the PC and the Internet were invented, including commercial virtualization and nanotechnology invented in the 21st century.
Therefore, AI is now considered the 26th GPT technology, so AI should have a huge room for development. If we assume that the steam engine invented in the 18th century was the steam engine 1.0 era, the 1.0 era was to surpass the physical strength of human beings, but if we use an inaccurate analogy, we are now in the steam engine 2.0 era. What are its characteristics? It helps us humans to surpass the boundaries of our intelligence. Therefore, AI has a huge room for development.
Let's take a look at our current AI. In fact, from a marketing perspective, it is packaged as a kind of cognitive computing. What is its technology? It is deep learning, and its algorithm is convolutional neural network computing, etc.
The AI we are talking about now is just one of many AI categories, and it has achieved very good results in some fields. The scope of artificial intelligence is very broad, and it is not just the artificial intelligence we are talking about today. There are many, many artificial intelligence methods. I believe that artificial intelligence will have a lot of room for development in the future.
The artificial intelligence we imagine is at least composed of the ability to learn, the ability to reason, and the ability to make decisions, but frankly speaking, our current artificial intelligence is actually still an analysis of correlations after analyzing big data.
For example, the sun comes out and the rooster crows, which is a correlation. So can we infer that "the rooster crows and the sun comes out"? Definitely not. I believe that today's artificial intelligence will definitely not be so stupid as to think that the rooster crows and the sun comes out.
But if we talk about relevance, these two things are indeed related, which means that our artificial intelligence is currently at a very, very early stage. But despite being at an early stage, current narrow artificial intelligence already has a considerable space for application.
For example, the office issued a report with the conclusion that I personally agree with that in the next 20 years, although machines are unlikely to demonstrate intelligence equivalent to or exceeding that of humans, they are expected to continue to achieve and exceed human performance in an increasing number of tasks.
If we agree with this conclusion, we can have more discussions later.
First, AI will change the organizational and personnel structure of all industries. As you know, our current industry is a pyramid structure, with leaders and managers at the top and grassroots employees at the bottom. But now, according to statistics in 2018, at least 3 million people have already worked with AI, or their boss is a robot.
You can imagine that the boss of a Didi taxi driver is an algorithm. In many fields, his boss is already a robot. I don’t know whether the 3 million data are accurate, but at least you can feel that your supervisor may be called Dr. Robot.
Therefore, many grassroots employees must get used to the fact that their colleagues are robots, and there will be some changes in the organizational structure.
Second, jobs have changed. Many people are worried that with the advent of artificial intelligence, many jobs will be replaced. This is true. Certainly boring, repetitive jobs that people are unwilling to do can be replaced by AI, but a large number of creative jobs cannot be replaced by AI.
Therefore, future work will definitely be combined with AI. Some jobs will be replaced, some jobs cannot be replaced, and new jobs will be created.
So there is no need to worry that AI will lead to mass unemployment. Instead, it will create new jobs, just like the invention of the PC enhanced human capabilities but did not reduce people's work, it just meant that there was a conversion or transformation of jobs.
So, from the current AI perspective, AI can be used to improve internal management efficiency, such as document entry and customer service. Now many customer service staff have already used artificial intelligence. Although you have finally got through to the customer service call, it turns out that it is a robot talking to you, which sometimes makes you very annoyed, but it does improve efficiency.
At the same time, for example, the access control systems we use now, such as facial recognition, as well as the current smart cities, safe cities, etc., these fields can be widely used.
What entrepreneurs care about most is cost. If AI can greatly save your procurement costs, you can know all the procurement prices through AI, which ones are reasonable and which ones are unreasonable, because you have information asymmetry. So you think you got a good price, but after using AI technology, the procurement cost can be greatly reduced.
When people talk about autonomous driving, there is a hot topic that autonomous driving will come soon, and hundreds of companies are engaged in intelligent driving. But from our point of view, it is almost impossible for a car to reach L5, or it is unaffordable.
Our view is that cars and roads must be coordinated. Simply put, we need a smart road and smart cars, and they need to interact with each other. There must be communication between cars, between cars and roads, between cars and people, and between cars and the Internet. Only in this way can the problem of autonomous driving be solved.
Therefore, autonomous driving cannot be achieved simply by relying on cars. Therefore, reaching L3 is already quite good. L4 still requires more efforts, but L5 is almost impossible. We can communicate about it next time.
Why is 5G so important as a new technology facility? As we all know, the bandwidth of 5G can reach 10G, or even 20G. 5G is not designed only for mobile broadband. If it is for mobile broadband, the current 4G can already reach 300M or 400M bandwidth.
More importantly, in addition to broadband, the second is the Internet of Things, which can support 1 million connections per square kilometer. Everyone knows the characteristics of the future smart world:
-
Everything is sensed, sensors are everywhere
-
Everything is connected, all sensors must be connected
-
Intelligence of Everything
5G and other existing connection technologies are the infrastructure of the future smart world. So on the one hand, they can provide ultra-broadband, and 4K TV, 8K TV, especially AR and VR, can all be connected with 5G.
The third is its use in the industrial sector, because its latency can be as low as one millisecond. So in one network, it can support the Internet of Things (IoT), enterprise communications, Internet of Vehicles, and ultra-broadband unlimited access. These are multiple applications. Therefore, 5G is the infrastructure of the future smart society, and 5G will definitely adopt artificial intelligence methods.
I can share some data with you. First, the current maintenance cost is 3-4 times the equipment cost. So you must reduce the maintenance cost through artificial intelligence.
Secondly, today's networks are becoming more and more complex. As we all know, 70% of network failures are caused by human factors, when maintenance engineers accidentally paralyze the network.
Third, after using artificial intelligence, more than 50% of potential failures can be predicted, so artificial intelligence can be used to reduce the operation and maintenance costs in telecommunications networks or 5G networks, and to ensure the improvement of network quality.
For example, some operators have asked whether we can maintain more networks under the same circumstances, or reduce the number of maintenance personnel by 50% under the same network conditions. These are some applications of AI in the industry.
Of course, AI has many good aspects, but there are also several problems with it.
Problem 1: The computing power is too expensive and we cannot afford it. It looks beautiful but is too expensive to use.
Question 2: Without good data, we cannot have good AI. Therefore, data must be processed and data quality must be guaranteed. Without artificial intelligence, there is no intelligence. Why are AI experts so expensive and engineers hard to find? This is a job for experts, and AI has not been popularized.
Therefore, all data needs to be labeled, so there is no intelligence without artificial intelligence. It is different from the artificial intelligence we imagine, and it is real "artificial intelligence".
Problem 3: Training is too slow. Training takes days or even months, and a performance may take a few minutes or seconds, but the training time is very long.
Question 4: Application scenarios. As we all know, artificial intelligence must have models. After much effort, a model is debugged and tested or tested in a certain scenario. If the level is very high, it is considered a top student. However, in the actual application scenario, the effect is not so good. The precision and accuracy may be greatly reduced, by 10%-15%, etc. Therefore, the top students in the examination room, the poor students in the workplace, and especially in the application scenario (maybe a little too much), but at least not as good as imagined.
So how is Huawei's current AI layout in the field of AI? Device, edge, and cloud. As you all know, our Kirin 980 has been embedded with AI chips since 1997. The 980 is more powerful, so in smartphones, our 980 is the vocal of the intelligent society.
In the field of AI, we released 310 and 910 in November last year, a full range of chips for cloud, edge and end AI. The Kunpeng 920 we just released is a cloud CPU. From the Kirin 980 mobile phone chip to the edge artificial intelligence computing chip to cloud computing, this is our layout.
So our value proposition is that we are a platform, including the cloud. As you know, connection is actually also a platform. Where would the data come from without connection in the future smart society? So we need to connect many sensors, connection + platform + AI + ecology.
Therefore, AI still has some defects. One is that its scope is relatively narrow. When the rules are determined and the results are clear, it can still do a good job in getting from A to B.
The second point is that it is more like research than engineering. What does this mean? Research is a craft, which requires constant debugging. The debugging is adjusted to this model. Technology is suitable for this scenario. It is not an engineering realization based on science. It has certain uncertainties.
The third point is that it works, which means that you have to design the model first. After the model is debugged and you think it’s okay, you then find an application scenario. Whether the data makes AI work is a very important foundation.
Also, now we have finally found a good model, but it will deteriorate over time. For example, I am at an intersection and I use a camera to automatically change the traffic lights, but as time changes, the model is constantly adjusting. The biggest problem is opacity. You don't know what is inside. It is precisely because of opacity that it is possible to analyze that the crowing of a rooster has called out the sun (of course, this is impossible). In other words, it is precisely because of opacity that this situation may occur.
Also, it is not 100% accurate, it is only a maximum possible accuracy, so it has some uncertainty.
In the field of AI, there is currently a shortage of talent. As we all know, there are no big data experts now. Overnight, there are only artificial intelligence experts. Despite this, there is still a shortage of talent.
Secondly, AI definitely involves data security and privacy protection. So, in the field of values, we actually need to discuss AI more. Not only should we say that AI has good sides, but we should also protect personal data and privacy, etc. These inventions should all be areas for discussion.
Everyone thinks AI is great, I think so too, but have you ever thought that AI is easy to be interfered with, or easy to be deceived?
Let's look at a photo. The left one is the original photo, which is exactly the same as the photo in the middle. But in fact, the photo in the middle has the noise of the image on the right added to it. But to our human eyes, the two photos look exactly the same. But through the machine, it recognizes that the middle one is not a person, but a bookcase, but to our human eyes, the two photos are the same. That is, AI can be deceived.
It doesn't matter if a photo is deceived, but what if a car is deceived? Would you still dare to sit in it? Who says cars can't be deceived? Of course, one is deliberate deception, and the other is if there is a scene that is not trained during training, will it malfunction? It is possible.
Therefore, there is still a lot of uncertainty in autonomous driving, and this is a case in point.
On January 6, 2019, a Tesla car hit a roadside robot. This means that the robot was not necessarily deceived, but it may have not been trained for at least one scenario.
So, despite this, autonomous driving still has a very bright future. When people think of autonomous driving, they imagine themselves sitting in a car and running at a speed of 120 kilometers. It may be a bit difficult to imagine such a scene, but is your sweeping robot autonomous? If a tractor is equipped with autonomous driving software, can it plow the land 24 hours a day?
Therefore, autonomous driving can be widely used. Don’t think of autonomous driving as just sitting in a car and driving at a speed of 120 kilometers per hour. It can also be achieved in certain scenarios.
Therefore, Huawei's vision is to work with everyone to bring the digital world to everyone, every home, and every organization, and build a fully connected, intelligent world.
thank you all!
- END -
◆ ◆ ◆
Recommended Reading
Pinduoduo has a bug: 100 coupons are available for free, and the freeloaders are going crazy
A large number of foldable screen phones are coming, will you buy one?
Exclusive! Zhang Zhengyou succeeds Zhang Tong as director of Tencent AI Lab
Alibaba and Tencent "fight" in Hainan Free Trade Zone
Redmi's first new product after independence Lei Jun: "Friendly competitors should not use the term cost-effectiveness"
The "2018 AI Adaptive Education Industry Research Report" was released on January 14, 2019. The current early bird discount price is ¥599, and the original price will be restored to ¥699 on January 20. It is a must-read for educational technology researchers, entrepreneurs, education practitioners, and investors. Scan the QR code for details.
京公网安备 11010802033920号