Baidu Kaiyu: Data is a very good competition barrier,

Baidu Kaiyu: Data is a very good barrier to competition

Reprinted from "Big Data Digest" to quickly focus on

Yesterday, the Geek Park Innovation Conference opened in Beijing, the Baidu Deep Study Institute (IDL) Vice Dean Kaiyu delivered a speech, elaborated the big Data age, the artificial intelligence and the deep study development.

The following is a transcript of Kaiyu's speech:

Good morning, dear friends! Before my speech, let's show you an interesting video. We see this car driving into our meeting, is from my left side to the right side, there is no one, it self-balanced walking, looks a bit strange. Why do we make such an interesting bike? It's a real bike, and it's not the bike that you used to walk on. Why do this thing, this is related to my speech today, is the Baidu brain thinking of human-machine relations.

You might be wondering, what is AI? This is a very serious academic topic, today this venue I think we have a passion for science and technology, do not need to be so serious, we are like I am a technology fan, saw such a movie, this small robot, it has a very high intelligence. First, the small robot has the ability to perceive, he has eyes, can perceive the environment: which is the road, which is the obstacle.

On the basis of perception, it will think and understand, understand what? For example, this place is flat, it can go, this place is down the slope, it can not go, this is understanding. On the basis of understanding, it says it will control and make decisions. So you see, from perception, understanding to decision-making, in a small robot embodied in the most vividly, this is very intuitive to feel the aspects of AI involved.

The essence of intelligence is learning

In today's market, filled with smart devices, smart phones, smart watches, smart cups, the previous paragraph of Google said to do smart spoon, we are doing smart chopsticks. Why are they called smart products? Are they really smart? Let's think about that. We put all our products aside, and on the other side were the old radios of the 670 's. They are different, in the case of human control to complete a number of tasks, from this point of view, they do not seem to be much different.

Thinking about this problem can actually help us differentiate between these smart products and what might not be intelligent. For example, if this is the same age, I learned a text in elementary school, told the story of the bench, Einstein made a small bench, do very rough, according to this situation we can say that he is not smart. You can come to the conclusion that this person is not so smart, compared with the children of the same age and the small bench. Einstein before the bench, every day is doing a lot of bench, every time in ascension, according to his past experience and lessons to improve his next time. If you look at it dynamically, you will find that he is a constant learning curve of growth.

What is the essence of intelligence in this? Is that you are constantly evolving with experience, and as our theme is called Awakening Evolution, Evolution is actually the essence of intelligence. What is the experience in this? Experience is data. You may not be smart today, but you will become smarter tomorrow, the essence of intelligence, as you learn from the past and continue to improve. So learning is the essence of intelligence.

Empirical data and deep learning

There is a very important aspect of this, which is experience. In the field of computer science and artificial intelligence, we call it empirical data. Experience is data, and data is experience. This involves the so-called big data that we often mention today, because we are now ushered in a great era, which is the age of big data. Let's take a look at how the mobile Internet has brought us profound changes.

For example, this is the square in the doorway to St. Peter's Church in 2005, and many of the believers await the results of the pope's election. Eight years later, the new generation of the Pope re-elected, the same group of people, we see how the mobile Internet changes the world, everyone connected to his device, these devices are constantly producing data. Everyone here today, like me, is connected to an average of two or three devices, and I have two phones in my pocket.

But this data, by 2020, is showing a growing trend of explosive growth, even more exaggerated than the forecast. A while ago, Wuzhen General Assembly Masayoshi Son said, 2020 everyone to connect 1000 devices. This is not necessarily so exaggerated, each button may be an intelligent connection device. In the era of mobile internet, everything is connected with the explosion of data. A lot of data gives us unprecedented opportunities, so we're connected from everything to the big data age, and the big data era comes with ubiquitous intelligence.

In recent years, we have often heard in the field of artificial intelligence that the most striking area is deep learning. With the big data era is quietly happening, which has its profound background of the times. Let's look at this curve, our horizontal axis is the data scale, the traditional artificial intelligence algorithm, the data to a certain extent effect no longer improve. For deep learning, we see what the curve is. As data continues to grow, the effect continues to increase. So deep learning can help us to absorb the dividends of the data. From a technical point of view, if you look from a business perspective, you will see if your business model is also Chen Xingguo Ian's curve as data grows, what does that mean? I want to congratulate you, you have a very good barrier, a competitive barrier.

Why? Because collecting data takes time, it takes a lot more time for others to get the same ability, so you always run ahead, and search engines are a pattern. The business model based on big data is a good barrier to competition.

The landing point of Baidu Deep learning

Baidu in the past two years in the field of deep learning can be said to have done a lot of work. To this day, Baidu brain This project has made considerable progress. It is a system of learning, evolution, and evolution without any time. Two months ago when I introduced to you, I introduced the company aims Yangzhou parameters to build the world's largest deep neural network, today I changed company aims Yangzhou to according, can build a larger depth of neural network.

What are the specific landing? Deep learning is inspired by the structure of the brain. Today we are more from the statistical and computational perspective of this problem, there are many scholars from the biological perspective of the simulation of the brain to do this research, the two roads should go hand in hand, and help each other.

As for the deep learning of Baidu in the interior where the landing, we can see in search, advertising, image recognition and speech recognition in these areas, there are considerable progress. For example, we look at the search, no use of deep learning, green is our search satisfaction above the leading competitor ratio, with deep learning, in the past year, the proportion of our ascension is such. From another perspective, it is also one of the most successful examples of deep learning techniques currently used in natural languages.

Of course, in the field of advertising, the entire Baidu Phoenix Nest advertising system has entered the era of deep learning, advertising CTR Increase is very significant, the specific number is a trade secret, I can not speak, is the percentage of double-digit promotion, the company's revenue is a huge contribution.

In the voice of this piece, we can say that in the past two years, the voice recognition rate is increasing, mobile search Chinese speech recognition We are the world's leading. In the field of images, last year we had a single model on the object detection model that was the best in the world and Google was second.

You may be concerned that in the past few years, the development of technology in the field of image classification is a process of quantitative change to qualitative change, we still look at the image classification example. Here I show you how the best grades in the world have changed in recent years. 2010 the world's best results in that place, the whole deep learning goes forward, and we see that every year it evolves at a very fast pace. At the end of 2014, this result is to 5.9% such error rate, this is Baidu to do, is also the world's current highest image recognition accuracy, we may be very concerned about, it than human ability is what level. The level of human beings in this place, nearly 5% of the error rate.

We just talked about the image and voice, reflected in our products is Baidu translation app, you can speak in Chinese voice, immediately automatically translated into English. It can also take pictures, foreigners come to Shanghai, see such a menu, pictures after the image can be recognized translated into English, you can understand what the dish is.

With artificial intelligence technology can help children have a happier childhood, summer vacation can have more time to play. This is not a direct help for you to do the problem, it is to tell you where the knowledge points.

As for the technology development of face recognition, we also report the latest achievements. On the most common face recognition test Facebook announced that the best results were 97.35%, Baidu colleagues made more exciting progress, the best facial recognition accuracy is 99.62%. Researchers at Stanford University recently made a test, with the figure reaching 99.2% per cent.

I have to clarify that the machine algorithm to achieve close to the level of human, in fact, just say the data level above the level of human beings, overall speaking, now the level of machine intelligence away from people or there is a gap, we are still in the effort.

As a child, you may have done a look at the pictures, you see the picture to use natural language to describe the picture of the story is what. For the left side of the picture, if a natural language description is a yellow bus, for another image we use the natural language of the sentence "Afternoon sunshine into the room" to describe such a scene. These languages are not actually written by humans, but are machine generated.

Many artificial intelligence algorithms are constantly accumulating its potential energy, changing all aspects of people's lives, such as the field of robotics, the specific example is highly automated driving, this is the team of Baidu Deep Learning Institute of the various cars in the conversion, this is one of our cars above the Perceptron demonstration, integrated radar, infrared, Camera, each perceptron covers different aspects, has its own advantages and disadvantages, how to integrate together on the situation of the road to have an understanding, automatically go to drive.

We have a three-dimensional understanding of the road based on these sensors, which is more accurate than people. For example, today's navigation system only knows which road you are on, and in which lane, we use a more advanced sensing system to position the car on the driveway with a precision of 10-20 centimeters. We can foresee in the next 1-2 years, the navigation system will be a tremendous change, in the streets of Beijing you have to go to the overpass, sometimes there are 35 points in front of the fork, how to go you will be blindfolded, immediately technology will help you.

This is our needle system, with deep learning to develop image recognition technology, with low-cost cameras to help you complete the task. No one's automatic driving is not the goal we pursue, it is difficult to do, ten years later may be able to do, the middle is the gradual realization of the process, it starts from the auxiliary driving, through the sensor to inform the driver what happened. The whole system is for people to enjoy the convenience and security of the transportation system, but not to replace people. We emphasize the driving of the people in the car, this is back to the beginning of my speech topic-on the human-machine relationship thinking.

Great technology is great, it is to make every ordinary person more great

Let's review a little bit about the origin of AI development, which is largely due to Turing testing, where artificial intelligence is a machine that can replace people, and it's not good for us to have some of our own thinking. We in Baidu used to do products inside, you can reflect a little bit, there are two products. The first is the PK Big coffee, an ordinary person to take a picture of themselves, find a very similar star, users are very welcome, every ordinary people have to do star dream. Another product is to take pictures of poetry, image recognition is not only recognition, it also generates a poem, more complex and advanced. But from the user's reflection, the user's interest in this photo-writing poem is not so great, why? Because this product is just to show how great the machine is, what it has to do with every ordinary user.

Let's look at another example, on the left is the milestone in the history of Ai--IBM's deep blue, the ability to do chess can reach the level of human champions, behind IBM nothing happened. The other side is the artificial intelligence technology, is the search engine. Search engine makes a child in Beijing, with the remote village children as long as there is network, the ability to acquire knowledge is the same, a child in the village can also acquire knowledge, can have a dream. This technology actually has a huge impact on society.

We go back to the beginning of the video, this bike can follow you. When you run, clothes can be put on top, water can be put on top, you can easily run. Or when you blow a whistle, it can follow you into the room. What do you think, this is a car? I think the answer here is the same as I thought, it's like a little black foal. So that's one of our thoughts on human-machine relations.

We believe that great technology is great, not to make the machine greater, its most important meaning is to make every ordinary person greater, more creative, his life is better.

Finally, I revise a quote from Chairman Mao: The World is ours, it is also a robot. But in the final analysis it belongs to the people who can create robots that can live in harmony with robots and use robots to serve me, thank you.

Baidu Kaiyu: Data is a very good competition barrier,