Depth | Kaiyu: The road of autonomous driving based on deep learning

The 2016 is a very important historical node, signifying that the AI system of unity of knowledge and line will go to the historical stage. It changes not only the next go, it will change a lot of things. --Kaiyu

On the "Adas and autonomous Driving Trends forum" of the "2016 Smart cars and Shanghai Forum", Dr. Kaiyu, founder and CEO of Horizon Robotics, delivered a keynote speech entitled "The road to autonomous driving based on deep learning".

There are technology popularization, industry observation, and a good outlook for future life. Everything you want to know about deep learning and autonomous driving, we're telling you today.

1 Deep Learning · Birth and growth

Everyone is talking about big data, like every middle school student is talking about sex, but they've never experienced it.

Deep learning has been proposed since 1957 to simulate a single neuron through perception, and now spans more than 60 years. For a single neuron, we extend it from time, space latitude, and the latitude of the correlation between them, to construct various complex neural networks, and then to do a lot of things, and this innovation continues.

Why should deep learning be valued? I share my four views on many occasions. The first point is that deep neural networks do play a role in structure and behavior, such as the neural networks we speak of Alphago, which are really affected by the visual neural system. I think from the point of view of engineering and application, in fact the larger reason is mainly the latter, one is particularly suitable for big data, one is "end to end" learning, the fourth is flexible modeling.

Why is big data appropriate? If many of the traditional intelligent algorithms for various reasons, such as the reasons for computing complexity, statistical reasons, the data scale to a certain time no longer grow, but its effect as the scale of data growth to a certain extent will grow. For deep learning, we see in many areas that continuous data growth continues to improve. So, deep learning is a very interesting relationship with the age of big data.

The ecology of today's technological innovation is that everyone is talking about deep learning, including a lot of creative companies. It reminds me of a joke about three years ago, about big data. Today each of us is talking about big data, just as every middle school student is talking about "sex", but they have never experienced it. In fact, today's deep learning is a bit of this flavor. First, the real it needs so powerful a calculator that requires so much data. Second, it is not simply to ask to take out some of the system to take out the use of a good, deep learning to solve your particular problem. Need to have this ability, not only to innovate once, but to continue to innovate, and finally really have such ability. The practice of companies or colleges in this area is actually very small.

2 Deep Learning · Wise vs Change Home

If it's just a perceptual level thing, it's a wise man who sits quietly over there. Decisions must be made and actions must be made.

From 2006 to 2016, the development of deep learning, in fact, there are mainly three driving force. The first driving force is big data-from the Internet to the mobile Internet, the massive amount of data generated. The second driving force is the surging calculations provided by semiconductor companies that allow us to deal with these massive amounts of data. The third driving force is the model and algorithm, from the very beginning of the simple structure of the deep neural network to some of the progress being made today. In fact, this progress has not stopped. Many of the latest developments that are underway may be more exciting than what I saw ten years ago. So it's not slowing down, it's speeding up the process of running.

The most notable 2016 is the latest advances in neural depth networks represented by Alphago. We talk about the past decade, whether it's convolutional neural networks or other neural networks, the thing that actually does is perception, which listens to the contents of our voice, to see what is in the image. But if it is only a perceptual level of things, at most is quietly sitting there with a wise man, and did not change the world. How do you change the world? Decisions must be made and actions must be made.

Reinforcement learning is actually such a machine learning framework, what is it? It is the game that describes a problem as a decision-making system and the environment, and if it takes action it will change the world. The world will give it a feedback, is a carrot or a stick, say you do well or not good. At the same time the world will change, to convey to it a state of change. If we describe this simple difference framework, how do we make a series of decisions to optimize a final goal? The ultimate goal, for example, is the return on investment to the end, a series of choices is whether to buy or sell today, how much to buy; speaking go, the ultimate optimization is how you win this game, a series of constantly to play chess to make your decision; in smart driving, making a series of decisions is accelerating or slowing down, is left or right, every time because of your change, The car around you will see you change and change, so this status has an update. Finally, a series of decision-making optimization, the result is from a to B to the safe, efficient to reach the destination.

What are the challenges of deep learning so beautiful and beautiful? Identify the car, identify the person, can identify the monkeys on the streets of India. Computing platform, including two aspects, one is the cloud computing platform, but also in the front-end controllability, how efficient, how low cost, how to do real-time. The third is system integration, including different sensors, including connections to the entire control system, which is a highly complex system.

One of the issues that people don't care about is controllability. Deep neural network, very efficient, but there is a huge problem is that if something went wrong you do not know what to do, it seems to be a black box system. This question is very important, I remember in 2005 when the development of airbag triggering algorithm in Siemens, we made a more advanced algorithm, in fact, if the accident occurred, we can conclude that it can save 30% of lives, but the final system is not adopted, the Product department is not adopted, Because it is a relatively black-box system. But if there is an accident in court can not provide a transparent explanation, what is the problem, this is the 1th. The 2nd is that it does not know how to improve after the problem, so this thing is beyond the technology of a problem, the problem is very important, if it is not resolved, it can cause a lot of problems. This is a traditional depot will be very concerned about, but many of our algorithms are only to pursue accuracy, in this respect is not necessarily fully recognized.

3 Autonomous Driving · Build a car for a horse

Imagine you're riding a horse. In a steeds scene, horse control at the micro level can do better than people, but you need to control that horse at any moment.

There is no doubt that autonomous driving faces a complicated road condition, especially how to adapt to China's traffic conditions? I used to inspire my colleagues, I said that as long as at the five crossing, if you can solve the problem of automatic driving, then this technology is the world. People who have been to Beijing may know where the five crossing is. In Europe, the United States has solved the problem of automatic driving, in the world's largest car market does not necessarily work. Of course, on the real side, you need to solve a lot of problems, including cost, reliability, ethics, security and so on.

The first concern is undoubtedly the architecture of the autonomous driving system that Google and Baidu have taken. This architecture actually has a feature, which is to run from the start to unmanned, hoping to cross a lot of intermediate steps. Based on such a feature, the high-precision map plays a very important role at this time. The core is that unmanned driving is actually going along a path, which is relatively not too difficult. What's the hard part? Is the shift, the difficulty is from the main road to the Auxiliary road, these things need the car to know exactly where it is. Therefore, the high-precision map will be the future of the construction of unmanned or highly automated driving technology a basic facility. But the problem is that we are still far from the true height of autonomous driving or unmanned driving.

With regard to Adas, I never called unmanned driving, because I felt that driverless could be a very vain piaomiao thing. From the industrial development, technology development path I think it should be more rigorous. Like Google, there is unlimited resources to do, hoping to achieve a road of unmanned driving, I think not only in the technical challenges, in the commercial also lack of feasibility. They may be reflecting on this issue recently, and they must be moving forward in a step-by-step way.

I was in the media interview almost a year ago, in fact there is a metaphor, in the future we need to pursue this goal of automatic driving, it is actually building a human car relationship, like today's relationship with the horse, the horse on the road will be fully aware of the environment, will be judged in a very timely and efficient manner. We can even imagine you riding a horse. If you are in a steeds scene, horse control at the micro level can do better than anyone, but you need to control that horse at any moment. So I think that the future of your driving experience, from the point of view of autonomous driving, may be almost as much a feeling as riding a horse.

4 Skyline · Define everything Intelligence

The internet thing is the internet company, outside the Internet physics Company we have to do something interesting. Can touch the physical world, how we make them intelligent.

What is the horizon doing? The mission of the horizon is "Define the brain of things", and now my interest and mission is that the internet thing belongs to the Internet company, the physical company outside the Internet we have to do something interesting. These things are around us and are the things we can touch in the physical world, how we make them intelligent. One of the most important categories is the car. How to construct a brain platform, which is first a software, is a deep neural network based operating system. Because I think from today on we need to focus on, in all of these hardware, not just software, we are not only to make it networked, first of all, the construction of a software operating system, personally think that the operating system is based on deep neural network of an operating system. Second, the architecture of the chip, which supports deep neural networks at the bottom, should be completely different from the design of many chips in the past.

Our main focus is on two scenes, one smart Car, the other smart Home, along these two lines continue to evolve forward. I just talked about the idea that we need the architecture of a new deep neural network chip, and I think that this understanding, whether it is the technology development of us or other companies, actually supports this view. I think we are not only the future to make it more powerful computational power, but at the same time the depth of the neural network itself algorithm logic and chip design will be common progress forward. For example, we say (Tianhe second) is the world's largest supercomputer, today its computational power is undoubtedly, but it requires tens of millions of watts of power consumption, but the human brain its calculation is actually the same as (Tianhe second), which is not only a physical, principle-based power consumption design, in fact, also includes the design of logic algorithms. So recent deep neural network in the algorithm level of a hot topic, in fact, how to design a very low-power hardware, can realize the deep neural network structure. It is not necessarily a highly integrated chip that can achieve very complex deep neural networks.

What is the horizon looking at now? What is the focus of the horizon's design depth neural network? How do we design this deep neural network to support the high-precision computing requirements required for autonomous driving in the amount of computing resources that can be purchased within RMB 200 yuan? So what I'm concerned about is what the 200 or 150 yuan can buy today, and I'm just going to develop a deep neural network algorithm for it, and we'll develop it specifically for the next 150 dollars of computing resources that we can buy. So we must pay attention to this trend. Is the computing resources between 100 and 200, and we're going to do the best with this algorithm.

Transferred from: https://mp.weixin.qq.com/s?__biz=MzA5MjM0MDQ1NA==&mid=402209186&idx=1&sn= ffb67895bed3a825558cfa7f749b9f0b&scene=1&srcid=0325bbxfybijapnsocpmkffz&pass_ticket= Jgroxcx0dyyu8qipzcgsijoyhhoyud%2f4rmurdbq96pegyem6suu1nzd0ilsjbz22#rd

Depth | Kaiyu: The road of autonomous driving based on deep learning