Li Feifei, the first generation of computer vision scholar

makes it difficult for machines to understand the world. A 2-year-old child can recognize “cat” from the picture after seeing the cat once, but the computer can’t.

and

Li Feifei are the first generation of scholars who use deep learning to solve computer vision problems. Inspired by 2-year-old children’s map recognition, they use big data to train machines.

,

Chinese, women, artificial intelligence, scientists, “washing sister”, “cleaner” and “counter attack” together add a touch of legend to her life.

on March 8, 2017 local time, Google held Google cloud next conference in San Francisco, USA. Li Feifei made a speech at the meeting.

and

artificial intelligence is a good story, the most important material is talent. Technology giants are doing everything they can.

and

Google tried their best to attract Geoffrey Hinton, the father of “deep learning”, Yann Lecun, a leading scholar of deep learning in Facebook, Russ salakhutdinov, a star disciple of Geoffrey Hinton in apple, and Wu Enda from Baidu As the competition for talents becomes more and more fierce, the scientists who once buried their heads in university labs have set foot in the industry to help this wave of AI.

and

have been around for nearly half a year. The most famous scientist came out of the mountain. Li Feifei, another Chinese scholar who is as famous as Wu Enda, joined the Google cloud department as the chief scientist of the Google cloud AI. In the world of AI, Li Feifei’s fame is like a thunderbolt. She was the director of Stanford Vision Laboratory and Toyota Stanford artificial intelligence research center. She was born in Beijing in 1976, and she was the youngest tenured professor at that time when she was 33 years old.

,

and “laundrysister” are the most important figures in the game of go. She has been focusing on computer vision research for 15 years and has been appraised as “the person who changes the direction of image recognition in the field of computer vision”.

and

are so-called computer vision, which teach computers to understand the world. This is the goal of AI research for a long time. However, for people, a rather simple cognitive process is very difficult for computers. One of Li Feifei’s most important achievements is that he launched Imagenet in 2007, now the largest image recognition database in the world.

,

are Chinese, female, artificial intelligence and scientists. Together, these labels make people look at Li Feifei with great admiration. In particular, “washing sister”, “cleaner” and “counter attack” always appear in her relevant reports, adding a touch of legend to her life.

and

, born in Beijing, grew up in Sichuan Province. At the age of 16, she immigrated to Parsippany Town, New Jersey with her parents. At that time, the whole family had poor English, and their parents didn’t have a good source of income. They could only do some jobs like supermarket cashier and camera repair to earn a small income and live on the edge. Two years later, Li Feifei got a full scholarship from Princeton University. This made her famous in the small town. The local newspaper also published her report titled “the American dream has come true”! 》。 During the period of

,

, ,

and

University, Li Feifei borrowed money to buy a laundry for his parents to run with his keen business sense. From Monday to Friday, she had classes in University. At the end of the week, she went out of the lab and became a “laundry girl”. “I love Princeton very much,” she once said with a smile, “but I also love my laundry very much. Without any of them, I am not what I am now.”

and

Li Feifei liked mathematics and science very much since he was young, so he chose physics as his major. “At that time, with the dream of becoming Einstein, physics can be said to be the most basic science in the human world. But in the process of learning physics, I found that at the beginning of the 20th century, the greatest physicists, including Einstein, began to think about problems that had shifted from physics to biology. They are thinking about where people come from and where people’s wisdom comes from, so I began to pay attention. I also pay close attention to neurobiology. I did several summer internships in neurobiology in college, and I thought it was really fun. ” It can be said that at this time, the interest in human intelligence and biology laid a trail for Li Feifei to enter the world of artificial intelligence later.

and

were in the bull market when she graduated in 1999, and wall street was booming. Many famous investment banks and consulting companies threw olive branches at her, but she refused. Instead, she followed her heart and went to Tibet to study Tibetan medicine for a year. Her passion for biology never stopped. When Tibet returned, she went to California Institute of technology to study for her Ph.D. and chose the direction of cognitive neurobiology and artificial intelligence. Later, she “mistakenly embarked on the path of artificial intelligence”. Li Feifei, the first generation of

and

scholars who solve computer vision problems, is the first generation of scholars who use deep learning to solve computer vision problems. It’s hard for machines to understand the world. A 2-year-old can recognize “cat” from a picture after seeing a cat once, but computers can’t.

“we use many machine learning methods of probability, but we have to design it by hand, its ability is very limited, and all data sets are very small. By 2007, I had just become a young professor myself, and I was thinking about how to break through this bottleneck. ”

and

at that time, Li Feifei had become a professor in the computer department of Stanford University, but the field of image recognition was still cold. Her colleagues advised her to change her direction in order to get a lifelong teaching post, but she did not listen.

and

one day, she suddenly realized that because the human eye can get an image every 200 milliseconds, regardless of the two-year-old children can recognize objects, but he has seen hundreds of millions of pictures from 0 to 2 years old, which is the geometric multiple of the computer, because the human eye is always absorbing the images in the natural environment, according to which, Li Feifei proposed the concept of big data.

and

she began to grab photos from the Internet, label them, and train the computer for in-depth learning, that is, to give some algorithms for the computer to learn by itselfLearning to recognize. Li Feifei and his doctoral students downloaded hundreds of millions of pictures from the Internet. It will take 20 years to complete the manual annotation without eating, drinking or sleeping.

and

were rescued by Amazon’s crowdsourcing platform, on which she hired netizens from all over the world to tag the pictures. In the meantime, the laboratory was short of staff and could not apply for funds. At the most difficult time, she even wanted to reopen the laundry and raise money for the experiment. Finally, 50000 netizens from 167 countries spent three years to complete the tagging of a large number of images, and then the landmark Imagenet was born. The important value of

and

Imagenet database lies in that it is open source and can be accessed and used by every laboratory. According to Imagenet, Li Feifei launched the annual visual recognition challenge, inviting technology giants such as Google and Microsoft to participate, and promoting the communication in the field of image recognition and artificial intelligence.

and

Imagenet are equivalent to an algorithm test room. Participants can use its huge question bank for the test. The higher the accuracy, the better the image recognition algorithm of the participants. In order to compete for the first place, the major technology giants have also devoted themselves to the research of image recognition. In the past few years, the image recognition ability of machines has been greatly improved, and the error rate is only about 5% (lower than that of human eyes). It can be said that the competition stimulates the development of image recognition, and Li Feifei is one of the biggest contributors to the development of computer vision and even the whole artificial intelligence.

and

earlier this year, Li Feifei came to Beijing to participate in the activities. In the interview, she will start with sentences like “to be honest, I don’t know alphago very well”, “I haven’t read the western world”, “I haven’t read the strongest brain”, “I haven’t read the book out of control”.

and

Li Feifei admitted that she is not a person who pays special attention to hot spots. “My personal experience is that the front of the eyes should be relatively open. If the front of your eyes is busy, this is not the best research direction. And open places are not hot spots, so you have to find your own focus. If I focus on hot spots, I don’t have today’s Imagenet. ” However, it is a lot of scientists like Li Feifei who have been studying on the bench for decades that have become the focus of artificial intelligence today.

and

joined Google this time. Li Feifei did not leave Stanford. She used her two-year academic vacation. In American universities, professors can enjoy a one-year vacation every seven years. From quietly burying her head in the laboratory to moving to the industrial world, Li Feifei said that she hoped to “democratize AI technology”, that is, let more people benefit from technological progress. So why Google? “Google cloud itself and Google’s data are very helpful for AI technology,” she replied

,

Yang Lan x Li Feifei’s dialogue

,

Q: Yang Lan

,

A: Li Feifei

,

, ,

and

. In the documentary “exploring artificial intelligence”, Li Feifei and Yang Lan talked about the topic of artificial intelligence. Yang Lan and Li Feifei are in the Artificial Intelligence Laboratory of Stanford University. Does the so-called deep learning of

and

include researchers’ training of machines? How to achieve it?

and

Li Feifei: the earliest in-depth learning is through our supervision training. We have the correct answer for each marked picture, and then we give it to the neural network. For example, the picture is pixel (Google mobile phone). It starts to collect some patterns of pixel.

and

then it puts it into a mathematical model, and then it gets the first answer, because it knows that the correct answer is a cat, then it finds out that it is not right, then we will correct some parameters again through the difference between the correct answer and the wrong answer, and then repeat like this. Thousands of times later it gets the right answer.

and

if you see a picture, the protagonist is a cat, its background is a used goods market, or the cat is chasing another cat. Does the machine now have the ability to distinguish between background and protagonist, or the relationship between them?

and

Li Feifei: now, we have developed to object in complex background, as long as it has a reasonable size can be recognized. The next step is not only to identify a cat in this picture, it may have a messy background, where is the cat, we can frame it out, which has been achieved. When a cat chases a dog or another cat, we begin to do these works. A paper has just been published in the laboratory, which is about identifying movement and relationship, but it has not been completely achieved, such as what expression a cat is, where they may go, and we have not yet done more.

and

I read in an article that it’s very common sense for us, such as a cup of water on a table, and for a machine, it’s difficult to distinguish the relationship between the cup and the table. Do they grow together?

Li Feifei: so our AI scientists often don’t worry about AI becoming the terminator, because they don’t even have this basic knowledge. Although AI is very powerful now, it still stays at the stage where you say I recite big data. It remembers and remembers the big data. The storage and calculation are also very large. You Give it some pictures. It can recognize.

and

but once they enter the abstraction and enter the basic cognition of the world, including the relationship between gravity, the relationship between the cup and the table, when you lift a cup, AI can’t tell you that the water will flow back now, unless it has seen countless pictures, so it really has a lot to do.

and

you mentioned artificial intelligence with humanistic care. I agree with you very much. In the end, no matter who is studying artificial intelligence, who is using it, what artificial intelligence is doing for people, in fact, it has always been the interaction between people and computers. I heard that you are also paying special attention to the development of artificial intelligence in the field of escort. Can you introduce some information about it?

and

Li Feifei: actually, it’s related to my personal experience. My grandma has already reached 9At the age of 5, I am far away from her, and some of my family are far away from her. We are particularly concerned about her daily life. So about three years ago, I began to think that AI is really moving towards application scenarios, which will have a profound impact on human society.

and

I think about medical treatment, and I am not the first, nor the last. First of all, human health is particularly important. Second, medical treatment is particularly expensive. Third, medical treatment is a special data-based problem. Artificial intelligence is very good at data problems, and there is also an aging problem. So I started to cooperate with Stanford medical college. We did three projects to show that AI technology, computer vision and machine learning technology can make some breakthroughs in three different scenarios. The first scene of

and

is the intensive care room. There are many jobs in the ICU. In the United States, one or two patients usually have a nurse, and the nurses and doctors are on duty for 12 hours. 1% of the whole GDP of the United States is used in the ICU. If anything goes wrong in the ICU, it’s a matter of life and death, so we’ll work with Stanford University of science and medicine to help nurses and doctors through modern sensors.

,

MIT’s AI Lab and Stanford’s AI Lab are led by two women respectively, but what is the proportion of female scientists in the whole industry?

and

Li Feifei: very few laboratories led by Daniela Rus and I are historical coincidence, not normal. At Stanford Artificial Intelligence Lab, I am the only female professor so far. We have more than 20 professors. At Stanford’s entire school of engineering, fewer than 15% of women professors. In the whole field of artificial intelligence, women may not be more than 10%.

and

do you think this phenomenon needs to be changed? Is it like some people say that girls are not naturally interested in science? Or is it not as good as boys?

and

Li Feifei: I think we need to change. It’s not just a female problem, it’s a human problem. Every science and technology represents our values. If we want science and technology to be a simple and friendly value that represents all human beings, it includes men, women, black, white and yellow people, including people in different fields.

and

I often tell my colleagues and students a very interesting thing, that is, if you search a very simple word grandma (grandma) in Google’s image search, you will find that the first page shows all white grannies, and you think, if an alien comes to our earth and wants to learn from us, what is grandma? It was the white granny that I found. This small example can tell you that if technology doesn’t introduce our values and concerns, it will inadvertently only represent the values and concerns of some people. So I always stressed that whether it’s women or other people from different backgrounds, we must participate in artificial intelligence and technology.

,

article / Wang Yue, Tang an

,

editor / Yunran

,

data provider / sunshine media group

,

(also pictures from the network)

Comments are closed, but trackbacks and pingbacks are open.