Intelligent networking is not only a technical problem, but also an economic problem.
Author | Berries and tinctures
Edit | Twilight
The 6th Global Conference on Artificial Intelligence and Robotics (GAIR 2021), hosted by Leifeng Network, was recently held in Shenzhen.
At the conference, Professor Yu Fei, IEEE/IET/EIC Fellow, shared a speech entitled "Interconnection: From Quality, Energy, Information to Intelligence".
Professor Yu fei is a Fellow of the Engineering Institute of Canada, an IEEE Fellow, an Institution of Engineering and Technology (IET) Fellow, an IEEE Distinguished Speaker, a Director of the IEEE Society for Teletechnology (2016-present), and a Vice President (2017-2019). Selected as a "Global High-Indexed Scientist" in the field of Clariva Computer Science for 3 consecutive years (2019-2021). Google Scholar 20,000+ citations, H-index88. His research interests include connected autonomous intelligence, blockchain, machine learning, autonomous driving, and wireless networks. Serves as an editor for several international journals. A number of scientific research achievements and papers have been awarded. In 2021, Professor Yu came to Shenzhen as the Executive Director of guangdong Provincial Laboratory for Artificial Intelligence and Digital Economy (Shenzhen) (Guangming Lab).
Professor Yu's main research interests include: connected autonomous vehicles (CAV), machine learning and artificial intelligence, blockchain and distributed ledger technology, wireless network physical systems and security and privacy in networks.
In order to present the professor's wonderful speech to everyone "originally", The new intelligent driving has made an edit that does not change the original meaning.
The title of my report today is relatively abstract, "Interconnection: From Quality, Energy, Information to Intelligence."
I will report "using the car as an example". The first part is background knowledge – connected and autonomous vehicles. The second is hierarchical design, cross-layer design, and cross-system design. The third is the method of artificial intelligence interconnection in information. The fourth is the theme of the report, from the perspective of networking, from the "large scale" to consider the network connection, can be divided into quality interconnection, energy interconnection, information interconnection and intelligent interconnection. The fifth part is a summary.
Connected and autonomous vehicles
The huge impact of autonomous driving is not limited to cars and roads, but also extraordinary for society as a whole. Every morning we may wonder whether I am driving the car or the car.
A few years ago, people were optimistic about the prospects of autonomous driving, why mention autonomous driving? When people talk about artificial intelligence, most of the application themes are "I won't have to drive myself in the future."
Show you two very interesting pictures.
First picture: 1900 America's Fifth Avenue Easter morning traffic, can you see the car? In 1900, photography technology was limited, and everyone may not be able to see clearly, but there was only one car, and the others were all horse-drawn carriages.
The second picture is 1913, 13 years later, it is also the same day in the morning of Easter on Fifth Avenue in New York, can you still see the carriage in the photo? No, it's all cars.
They use the contrast of the two diagrams to express the fact that when technology wants to abandon you, it doesn't even say hello. This analogy is used to illustrate that in the future, autonomous driving will also be as rapidly iterative as the previous car replacing the carriage. In previous years, these people have used financing and technology reports to convince investors that autonomous driving will soon be possible.
Although the ideal is very full, the reality is very cruel. You may have heard of various examples at home and abroad, especially man-made accidents in Tesla, UBER and some large factories, including the problem that Tesla cannot identify white objects that has attracted widespread attention, resulting in various accidents.
Waymo's CEO also poured a basin of water on everyone, Waymo is a subsidiary of Google's autonomous driving, so Waymo has a say in the field of autonomous driving, since 2009, Waymo's self-driving vehicles have run more than 20 million miles on real roads and 2 billion miles in virtual environments. But Waymo's CEO says self-driving isn't likely to appear on real traffic on a massive scale for decades. What's the problem?
He recently commented that Technology is really really hard, and technology is too difficult.
Elon Musk also famously commented in July 2021. People are asking him, you said earlier that fully autonomous driving will be achieved soon, when will it be realized? Musk then shoved the ball in front of engineers and scientists in academia and industry, saying, "It's not my problem, it's not that I can't do it, it's that the scientific community hasn't solved the problem of AI science." He pushed the "responsibility" off the scholars.
Therefore, the inability to do automatic driving has little to do with Tesla, and it is a problem of "us".
I as a scholar and engineer are actually relieved to see this sentence, from the perspective of consumers, everyone sees that the attention will fall on the point of "automatic driving will not be realized in the short term", but as a scholar and engineer, we see the opportunity, why? After all of it is done, there is no chance, and if it is not done well, we still have a chance to achieve it.
So I've been wondering what the hell is going on? Opinions vary. Essentially, information is very different from intelligence. What's the difference? Self-driving cars can generate 5T of data a day, and various sensors are producing a lot of data, such as cameras, GPS, LIDAR, etc. But for autonomous driving, this information is not the same as intelligence. Intelligence here I define it as "driving a car," like steering, braking, and throttle.
Hierarchical design, cross-layer design, cross-system design paradigm
I would like to briefly introduce our design approach during the Information Internet era. It's the infrastructure for autonomous driving in the communications network, and we generally use DSRC, or C-V2X.
What design paradigm do we use in the Internet of Information?
The earliest is a hierarchical optimization method, that is, each layer handles a separate task, such as the power of the physical layer, AMC (adaptive modulation and coding), the MAC layer handles different user ratings, RLC handles retransmission, non-retransmission, reliability; PDCP processes packet compression; RRC (Radio Resource allocation), CELL Selection, Handover, Optimizations such as Admission are at this level. This part of the applications can also be optimized, like which codec, H.261 or H.262 to use.
Users can also participate, which is how optimization was done decades ago, but hierarchical optimization does not meet the overall system requirements. Later, cross-layer optimization appeared, that is, the integration of layers to optimize. For example, the upper and lower layers are combined to optimize, and the effect is better.
For example, the application layer transmits real-time information or control information for automatic driving, which requires high latency. The physical layer has real-time network information, and joint optimization will produce good results. The next step is to optimize across systems, communications and networks become subsystems, and other subsystems are also important, such as computing systems, which consider edge computing, cloud computing, and material computing. The other part is storage, please don't ignore the control part, it is not what traditional communication and networking do, it belongs to other subsystems.
Here communication and network are grouped into a subsystem, and it is necessary to optimize them together. Why? Because of different applications. For example, autonomous driving, or the now more fiery metaversics, AR, VR, have a higher demand for computing, and only the network cannot meet the needs of the entire system. For multimedia transmission, there are more requirements in terms of storage and caching. I summarize it as a cross-system design.
In this regard, we review some of the work we have done.
Communication computing combines work;
A combination of communication, computation and caching;
Communication and control work together.
Each combination improves the performance of the network or system.
When we write articles, the performance improvement is particularly gratifying. But let's not forget the "bitter price" that comes with it. One of the trade-offs is a significant increase in complexity, from single-layer designs to cross-layer designs to cross-system designs, with more and more parameters to consider each time. A large number of parameters are optimized together, and although the performance of the system has improved, dimensionality disasters will also accompany it, which can also be called the first "curse".
Another problem is the Curse of Modeling, or "how to model." One-tier modeling is not complicated, but multi-tier modeling, cross-system modeling, and cross-network modeling are very cumbersome. Problems arise almost everywhere in modeling, and there's a famous saying: All Models Are Wrong.
So there are two curses: Curse of Dimensionality and Curse of Modeling.
AI Approach
Because modeling is becoming more and more difficult and complex, we hope to be able to solve network optimization problems in the form of artificial intelligence.
The theme of our conference is the "Global Conference on Artificial Intelligence and Robotics", and all of you are more or less people associated with artificial intelligence. AI is not a new concept, "artificial intelligence" was proposed in 1950, machine learning began in 1980, and it was until the emergence of deep learning (deep learning) with superior results in 2010.
Why can't artificial intelligence be mentioned at that time? Professor Hinton, the winner of the Turing Award, "fled" from the United States to Canada that year, also because of the "artificial intelligence winter". Professor Hinton has been researching machine learning and neural networks, and finally made a breakthrough in 2012. The development of artificial intelligence is not smooth, it can be said that there are several ups and downs.
We use Reinforcement learning more often. Machine learning can be divided into three categories: supervised learning, unsupervised learning, and reinforcement learning.
The article in the bottom left of the figure is the last chapter of my doctoral dissertation published in 2004. It is also the embodiment of a bitter lesson for me. At that time, it was impossible to publish the top article with machine learning or reinforcement learning, because the general environment did not approve. "Artificial intelligence" was a pejorative term at that time, considered garbage input, garbage output, can not produce Insight, failed to get the recognition of the industry and academia.
Reinforcement learning is an algorithm that I personally like very much, mainly because it can do a wide range of "movements", control the car, control the network, or control a wide variety of parameters.
The central idea of Deep Reinforcement Learning is simple: to simulate the process of animal or human interaction with the environment. Deep reinforcement learning can solve many "big problems," one of which is Alpha Go.
The core idea of Alpha Go is to use reinforcement learning to solve problems. Reinforcement learning technology has deep roots in Canada, and deep learning is led by Turing Award winner Professor Hinton. Reinforcement learning is led by another Canadian scholar, Richard Sutton.
Using the AI approach, using cross-layer design, layered design, cross-system design, is not without problems.
Data matters. The foundation on which modern artificial intelligence "thrives" is data-driven. Data-driven was a positive word a few years ago, it's not all from the model, but there is real data. But many small teams, small companies, etc. have relatively difficulty obtaining big data. Another form of data driven is interpreted: Big data leads to big intelligence, and Limited date leads to limited intelligence.
Returning to the topic of automatic driving, there are thousands of roads in the world, and humans cannot let the model specifically learn each intersection, each type of weather condition, and the driving conditions of each driver, and the lack of data is also an important factor in the current failure of autonomous driving to be commercialized on a large scale. Data-driven, so Limited data is Limited intelligence. Other challenges include data inefficiency, which requires a lot of data training. Poor generalization, weak generalization ability. Lack of interpretability, poor interpretability.
Something went wrong without knowing why. Can one party share it with others when they have large-scale network data? The yarn intelligence between machines requires a specific language and program to be carried out.
I think the current machine learning, artificial intelligence, is a bit like animal learning.
A 2019 Nature article mentioned that today's artificial intelligence may not be as good as animals. One example is the well-known saying we Chinese", "Dragons give birth to dragons, phoenixes give birth to phoenixes, and children of mice make holes." Describes the skills that already exist in the genes of living organisms. As in the text: Learning is NOT very important. This article is tantamount to pouring cold water on me. We have also been studying artificial intelligence and machine learning for some time, and if AI is not as good as animals, it cannot be compared with humans.
Later I hope to get the answer from the book, to understand the fundamental difference between animals and people. In this regard, "A Brief History of Mankind" is very clear. There is a subversive point in A Brief History of Mankind: the main difference between humans and animals is "Gossip", or "gossip ability".
Why "gossip ability"? Humans can deliver information that doesn't really exist. Like "spreading" gossip in the company, WeChat, Keynote, and even today's sharing reports are a kind of "gossip ability". We can only talk to the same kind, we can't go to the forest and report to the animals.
The term "gossip power" is less formal, so I found another formal term—Collective Learning," and many historians in the Big History project also wonder: Why are people smarter than animals? Looking back at the evolution from the Big Bang to the present, the basic conclusion is Collaborative Learning.
The first steps of Collective Learning are all learned from data, and data-driven is also the basic idea of artificial intelligence and machine learning today.
After the second and third steps, machines and animals basically do not have the ability to store intelligence, only people can do it. We can write a book with a lot of ink, which is a special ability of human beings. Similar to the "gossip ability" in A Brief History of Mankind, people can believe in information that does not exist and share intelligence. The above two points are informally "Gossip" and formally "Collective Learning".
You may ask: Can artificial intelligence now have "gossip ability"? Or learn from other agents? It is very difficult at present. Because there is no Incentives, trust, language. How can these capabilities be achieved? That's our next theme, quality, energy, information, intelligence.
Connected: from quality, energy, information to intelligence
Albert Einstein said: You cannot solve a problem on the same level that it was created. You have to rise above it to the next level. I was deeply touched by this sentence. "When you have a problem, don't think about it on the same level as the problem, but think about it on another level" to generate new inspiration. This is also the theme of my report today, considering the network connection from the "large scale".
Many media reports, including Elon Musk, agree that the greatest invention in human history was the wheel. The essence of the wheel is the interconnection of mass, which can quickly and efficiently transfer something with mass between two points. Wheels contribute to the formation of transportation networks. The second important invention is the interconnection of energy, which was previously gasoline and now electric energy. The third important invention is the Internet that people enjoy, which is the interconnection of information. In short, the "car" has witnessed the development route of the entire technology of human history. From the initial quality of transport, followed by the use of energy, to the current networking of information.
From a development point of view, Abstraction improves layer by layer. We wonder, how will we develop in the future? After a long period of reflection during the pandemic, I wrote an article. When optimizing the network, the layered design, the cross-layer design, and the cross-system design, in the final analysis, we are doing one thing, passing on the information. This is also what we call the Internet of Information, which moves information from point A to point B.
Tesla invented alternating current, which can be understood as transmitting energy. At present, information development is overloaded, and overwhelming information is everywhere in our lives, and obtaining information is easy for modern people. So I deduce from the article that this is a lack of intelligence - how to use the information when it comes. For example, an autonomous vehicle can obtain 5T of data every day, holding a huge amount of data, but the driving skills are still not high.
Can move intelligence be implemented at this time? Not yet. But we can consider using Blockchain (blockchain) to solve the relevant Wentong. The Harvard Business Review once published an article asking whether there have been things like Blockchain in history. The answer is yes, TCP/IP (Transmission Control Protocol/Internet Protocol).
First of all, they are all distributed, and the advantage of distributed is that it promotes innovation in a centralized manner, and it can also support other applications on a large scale, which is the similarity between TCP/IP and Blockchain. Blockchain is used in many ways, and our national reports often see Blockchain as an important technical tool for the digital economy.
We have recently published books on blockchain – Blockchain: Principles, Frameworks and Applications and Blockchain Technology and Applications; and the blockchain research site vDLT. There is also the problem of blockchain optimization, we mentioned that Blockchain is more in terms of implementation, less optimization.
The answer to why humans are smarter than animals points to one point – the Collective learning approach, which we've done recently to achieve through blocks.
Do intelligent networking from the perspective of the car. I have proposed a prototype in the article, which is also the prototype of the metaverse, and each car corresponds to the "digital twin" car in the metaverse and shares intelligence in it.
The other is an algorithmic innovation, Collective Reinforcement Learning, whose basic idea is also to imitate humans. Reinforcement learning was originally a single agent, and now multiple agents are fused, learned, and intelligent networked.
For the challenges of today and in the future. The challenge seems easy, just pass on the intelligence. However, in practice, it is frequently hindered. From the perspective of information theory, why can information be easily "moved around", and why can our Internet develop at a high speed? The key lies in the definition and description of the information. The current description of the entire Intelligence is very difficult. It is also a more difficult problem to overcome.
Future trend: intelligent interconnection
Internet of Information can be designed with layers, layers, and systems, and it has recently become more popular to use AI approaches. We believe that the Internet of Intelligence is the future trend, and it is even more important for machines and people to make the right decisions. Shared intelligence is not just a technical problem, it is also an economic problem.
Our Guangdong Provincial Laboratory for Artificial Intelligence and Digital Economy (Shenzhen) has just been established less than a month ago. Because it is located in Guangming District, Shenzhen, it is named Guangming Laboratory, which is one of the third batch of provincial laboratories approved by the Guangdong Provincial Government.
At present, we mainly focus on four aspects: one is blockchain and financial technology; the other is intelligent sensing and precision medicine; the third is machine learning and intelligent systems; and the fourth is ubiquitous perception and smart city. Thank you again and welcome to our laboratory.
END
"Unmanned" driving is encountering a "employment" shortage
Revolution in autonomous driving "perception"? Inspiration from the human brain and frog eyes