Original: Tan Jing
Linda Hua, the current head of MMLab. MMLab is the Chinese University of Hong Kong Media Lab, is Hong Kong Chinese - Shang Joint Laboratory. Head most of the time is spent in more laboratories, so Beijing's reporter wanted to interview face to face, not necessarily an easy task.
Finally, we met in artificial intelligence (AI) scenes of martial arts martial community together - the World Conference on Artificial Intelligence.
"The most influential global smart field scientists and entrepreneurs together in this land," the saying goes, it does not mention.
Meet the martial arts class martial head, the brain will "work hand Baoquan like" idea. Who knows, he expressed a few words of English mash, immediately pulled me back to the AI in the world. When Linda Hua unique Yuehua Jian always with a kind of professor (Professor) classes delicate and patience, if either ask for help, also its argument.
Practice superior martial arts, you need to enter the door, thanks to teacher. In academia, it is a relatively high altitude advantage, though not absolute advantage, however, the researchers located the platform can often play a decisive role. AI taught in many embrace the dream of the eyes of her students, MMLab is the door, Linda Hua is a teacher.
Today's teacher is also a student yesterday.
Time back to 2012, Linda Hua received his PhD in computer science from MIT. "Why did not stay in the United States?" He should not be the first time this question was asked, and he has made his own choice.
He smiled, and gave some details of the reasons.
He said: "Mainland China and Hong Kong have a good environment, adding MMLab can be quickly put to work in the Chinese University of Hong Kong and Tangxiao Ou teacher gave a lot of support."
Visible year, he chose to study platform when there is not much hesitation.
"What is studying in MIT greatest achievement?"
His answer was, come into contact with different scientific cultures, learn different research thinking.
He also stressed: "Innovative thinking is supposed to crash, I particularly value."
This is a point that needs to be paid special attention, and it is the only way to practice.
He told me, collision, resulting in a lot of innovations. This is not only his experience, but also Tangxiao Ou teacher of philosophy. Since important, he further explained the "collision."
He said: "Professor Tang also said before, new ideas and the exchange of ideas depends on the collision out of the collision of the lab innovation is important, the researchers stand in the forefront of the world.."
He recalled the time graduate student at the Chinese University of Hong Kong, talk to him and said:.. "Early face recognition technology has not been used in depth study after I learned at MIT, learning more partial statistics and probability modeling returned to Hong Kong Chinese when university professor, just living in depth study of the wave, the depth to do is to learn. "
A cycle, a researcher often the rest of my life, the opportunity for those who are prepared.
Linda Hua has a high starting point, but he continued to accumulate on this starting point, make every effort to research and academic students to a new height.
Five years passed and, from 2015 to 2019, MMLab scored a total of 99 CVPR, 38 Pian ECCV, 51 Pian ICCV, 9 Pian NIPS.
Today MMLab is no longer a martial arts martial, but belongs to the school of martial arts league.
2019 is a milestone.
This year, the Shang Dynasty with a number of science and technology and their world-renowned universities to build a joint laboratory, a total of 57 papers were ICCV. Count 62 of the same year CVPR papers received, a total of 119 papers selected will be the world's top two computer vision.
People often say that behind all the scientific research achievements, are studying hard, solid training. In fact, there are more important in the last part, which is the key more worth exploring - "? What is effective training."
With "Jiuyangzhenjing", we have to pay attention to how to correct the key to practice, obsessed by how to do?
"Do research (research) What is the most important thing is?"
Linda Hua said: where to find the real challenge.
Entered MMLab, Linda Hua hope that students, especially the first question is just entering the academic field of study students can gain a profound understanding is: What to do research (research) is the most important?
"The answer is not complicated." Linda Hua said.
"The most important thing is to do research, to find the real challenge lies. Many researchers patted his head in the lab, doing a paper. While this paper may be very successful, but not worth what the application, because the academic imagination questions and industrial landing problem to be solved, there is a big gap (the gap) between them. "
He paused, stressed:. "AI researchers in the process of contacting the ground to the real demand, which found academia did not notice a thing."
"MMLab students, do not send non-paper will be the top, do not send no breakthrough papers."
This sentence represents Linda Hua student expectations and requirements. He did not want students to have "another" academic thinking and habits in academic way.
In his world, outstanding and non-outstanding than two different standards, but in doing two very different things.
Papers have a decisive significance for the highly academic scholars. However, the cultivation of students from Linda Hua, from his understanding of academic education, we will not waste his sweat - just write "pretty" papers.
He wants a high-quality innovation, which is the motivation MMLab culture deeper.
One would like to beat the other, will become a martial arts master. Founded martial arts sects bent on thinking, will become martial arts master.
Just a little to know, we can know, the Chinese University of Hong Kong today MMLab never missing students, many solid foundation, good students come here especially.
Linda Hua described the students just move the lab, "A lot of students in the first year to MMLab, there is a certain stock of knowledge, but how to do research still in its infancy." In his eyes, each student's plasticity is very strong.
He admits: "MMLab expectations of students is to create a separate direction after graduation, with a team." For instance, he saw many of his students have graduated to the Shang Dynasty can directly lead (with) team.
"After he entered the laboratory, what kind of training and training students to accept?"
Perhaps the first time this question was asked, during an interview, Linda Hua Jingjing pondered for a while, he took out a "three-stage theory," so I had to quickly enter the state "record practice tips" in.
He stressed, MMLab not unique offerings Cheats culture, on the contrary, it is a researcher in the field of AI must experience three stages, the law also personnel training.
The first stage, know how to do a project (project), highlighted a "collar" word.
He will tell students what to do a project, which direction to explore, what technical route Yes. Students under his guidance, with the assistance of brothers, the gradual completion of a number of independent projects.
In the beginning of time, he will discuss "Professor guidelines and expectations" with students carefully. Linda Hua stressed that, in this process, will not force students do not want to do things. He said "absolutely not" the words when specifically added stress.
Because the students to do, though, is keeping professor guidelines, but students must put forward their own ideas, clear local interest.
He then according to the students' ideas around this problem is not really academically valuable, do not go in this direction will encounter problems of some fundamental barriers to and fro communication.
This process may be a month or longer. He believes that the process itself to create a sense of teaching.
He stressed: "The goal is to find a long-term professor led the students go on to do further academic direction." At first, he may give students more guidance, students observe and understand the situation of students step by step to learn to adapt. It will be under supervised circumstances, the gradual development of research students at this stage.
At this time, the second time Linda Hua stressed, MMLab students will not send no breakthrough papers. Because the target set low, is a waste of time for students.
In his view, ways of thinking and study habits is extremely important. If you locate the top will send a non-paper from the beginning, it would become "another" in the way of thinking, this way of thinking, not the laboratory culture system. This is from the perspective of the needs of students once again to interpret why not send non-top Papers.
The second stage, a prominent "independent" character.
Linda Hua said he would set a direction and students together, but there will be nuanced guidance (guidance). I am afraid that students need to find their own resources, most try to even the data sets are not.
In Linda Hua eyes, MMLab in a lot of games in the top, it's just exercise for students.
His confident and honestly say: "We have completely beyond the 'brush list the times', to exercise students' ability to solve problems with AI, long-distance running in the first phase of the development of the I gave them has been completed."
The second phase of the key tasks is to develop a direction.
"What we will discuss in this direction goal is probably even data sets are not, it would have to build their own, the algorithm to do it, experiment design, adhere to completion. This time, students need to develop a high level of project independently Ability."
The road of life is described in the "Research Explorer" an order and pick up on Linda Hua. How many martial arts person step by climbing the stone steps to "climbed steeply" below, looked up, four words awe-inspiring.
He continues, the third stage, a pass before graduation also highlight a "break" character. Students find their own research, to produce independent research ideas, adhere to in the end.
He re-emphasized the main points: "challenge themselves to find their own to find the problem."
"Get down to go through these three stages, the basic means graduation can work independently." This is Lin's teachings, he is also the wish of the heart.
"Students individual's situation will be different, and some partial-thinking, and some partial practice type, and some partial Engineering, I hope that after graduation, each student will study the formation of a unique personal characteristics of the path."
He adds.
The word "unique" is Linda Hua extra value. To some extent, the word contains a kind of "high quality innovative" gene. His speech revealed that students value the inherent characteristics.
"Whether students are preferences or preferences engineering research, you will find their place. After graduation, some students are willing to go to Shang technology, because today's technology has the Shang Dynasty was a large platform in computer vision. Some students are willing to go to the US for further studies . "
Linda Hua happy to see germination, jointing growth, the options of the future struggle for land is tropical rain forest, or plateau basin, he will not give restrictions.
His task is to train students to come out, and gene MMLab with.
From one culture to culture batch.
Professor Tangxiao Ou MMLab founded in 2001, ten years later, it has long nurtured a unique research culture. "How do we understand MMLab team culture?" Linda Hua replied, "Of course we have our own culture." But think for a moment.
He said: "This is my first time to sum up the cultural laboratory."
First, respect.
Respect students' innovative ideas. The emphasis here is not allocated research ideas, our role is to mentor (adviser), the focus of this role is to guide students to form the idea of the study.
Professor and do not touch the data and code in line, if just to find fault is likely to interfere with the students' creative thinking.
Students need to find the real value of their own challenges. When students develop the idea, he would throw the student first question - why did not solve this problem before?
Never expect him to tell the students, you have to do or not to do.
After a literature review of this problem may be done, it may not be thinking too clearly. Literature review to answer just one part of the problem.
To explain this critical issue, he immediately gave an example, like most of the classroom teacher asked students to answer.
The "timing algorithm, for example, students might say, ten seconds before the method subject to the restrictions of memory, several minutes or longer video analytics difficulties encountered, what different issue I want to study and previous studies of introduction class. "
Linda Hua said that the issue can not allow students to rigid answer, he asked the students will try to answer specifically the question of research papers and what are not the same as A, B and papers which are not the same. "
Second, value.
Assuming that this study has to do it, where the value of?
He stressed that is not limited to academic value, but rather to bring the value of human society.
"Or at the timing of a segmented network, for example, to solve this problem, it means expanding the length of AI processing time video capabilities, video previously not treated, can now be handled by technical means."
Specific studies he borrowed to explain the way of thinking.
"If you want to do something in the academic, first of all to give an unequivocal answer these two questions. If these two problems can be handled well, the problem to be studied actually been set up."
He stressed that an academic question, do not need to define a bunch of questions, one to two fundamental questions to be able to define the problem clearly. Adviser (mentor) in this role focus is to guide students to create valuable ideas.
Teaching and learning, A and ask, practice every day, every night to ponder the matter.
How MMLab scientific and technological cooperation with AI Shang unicorn?
"Before I answer, I must say, MMLab as research institutions, compared with commercial organizations, has a completely different mission. Mission determines the destination." He stressed the first sentence, he began to answer my question.
This represents, he was very clear understanding of the mission of research institutions of their own leadership is both cooked to think, want to have a thorough problem.
Linda Hua said: "Shang energy and a lot of different industries have different requirements for the contact partner AI, accumulated a lot of experience on the ground experience of these academic researchers invaluable.
MMLab will Shang Dynasty and Research Institute, the daily exchange very closely. Face practical problems, the first Academy of Shang Dynasty, the face more fundamental issue, the Institute will 'pass the torch' to the lab. "
Behind the "pass the torch" is full of confidence.
An action, two connotations.
Pass the torch had completed their task, arrived at their destination. Successor acceptable to the people not only work, but also trust.
"Pat on the back, man child, later on you."
Successor, turn.
The back is a responsibility and expectation.
"The fundamental problem in the form of research project management style to advance. Laboratory research process can not guarantee 100% solve the problem out."
When he expressed, look rational and firm, and gently shook his head, emphasized the "study of fate": "to do research because there is no 100% guarantee that no.."
. "But," he smiled, then said:. "In laboratory studies the issue of process, will put forward a lot of very valuable ideas Not only that, these ideas, the research team will make the experiment, prototyping come out."
Then, he directly expresses the relationship of the Shang Dynasty and MMLab of the face, without any intended meaning mildly.
"Shang Tang is a commercial organization, to consider the revenue, there may be some problems can not be delayed a year or so.
At this time, the advantages of the lab will have to play, because there are a lot of laboratory space to do it, will develop a more long-term, more innovative solution ideas. Shang Dynasty and MMLab have exchange mechanism, to ensure that 'find and solve problems of closed-loop' continuous cycle.
Laboratory innovative ideas become products in the shortest distance, and the market will give laboratories the most direct feedback. "
In Linda Hua seems, research voice calling for many years, however, the chain of cooperation among many agencies are not smooth, but Shang and MMLab natural and close cooperation.
Research funding, is a quite interesting problem.
Linda Hua clearly describes the two major sources of research funding, he said: "On the one hand, Chinese University of Hong Kong to support a large laboratory, professors do not need to worry about the cost, can be very concentrate on research.
On the other hand, Shang Dynasty Chinese University of Hong Kong Science and Technology and research investment agreement. In addition, a large Bay Area AI-related industry has developed rapidly, policy support. For laboratories, small pressure on costs, higher and wider field of vision, focus more on long-term focus on creative, better able to solve the problem of the industry completely. "
Homeopathy, he cited two examples:
First, how to deal with than the big label data a hundred times, thousand times the data?
Computer vision research will involve massive amounts of data, many of which did not play tag, and the depth of the traditional network is supervised learning.
So, how to deal with than the big label data a hundred times, thousand times the data?
This problem is referred to the MMLab.
As we all know, the quality and performance of the model and input data of a great relationship. Without effective clustering methods, so it will send data into the model quality. We need an effective, over hundreds of millions-level data clustering method.
MMLab recent papers are 2-3 on this aspect. An effective approach is to large-scale data clustering, in which each category as a different person after. But how in high quality and efficient mass data clustering is unresolved open issues.
This problem is not specific to a product upgrade, but the research team abstract solutions, to solve a fundamental problem.
The good news is that using a neural network clustering technology on all the Shang need to use huge amounts of data to enhance the performance of products has brought many business lines have used this solution.
Second, understand video work.
In 2013 and 2014, most of the work is handled short video to understand the video within 10 seconds, the classification of short videos with the way the machine is still very far from practical. Even after convolution neural network is proposed, because of hardware limitations, or no way to handle a long video.
A video MMLab traditional analysis, every five to take one. GPU can not put too much frame. However, the actual situation of the industry is the need to deal with a relatively long period of a few minutes of video. In this context, MMLab in 2016, proposed the partition convolution timing network.
"The timing of this method is no longer segment five, five fetch. Lab entire video semantically, every five divided into several paragraphs. Such solves two problems. Taken every five frames, high repeatability repeat computationally intensive. If only take a distant intervals, and becomes longer time scale.
When the team decided to take the frame toward the distant long effort, because the focus is to avoid duplicate calculations.
Ideas to solve the problem is not complex, but fundamentally the direction of thinking and everyone's different. Others are thinking about how to improve the network structure, but we changed the method of sampling.
This method allows laboratory has achieved year ActivityNet champion. It is also good news, since 2016, this paper presents a method had been used in a variety of video analysis Shang line of business products. "
Linda Hua would like to use these two examples to compare the difference between university laboratories and corporate research institute. University basic research laboratory space, once the basic research success, tremendous leverage benefits.
He has the ability of laboratories to solve problems of industrial sector more thoroughly, very confident.
School-enterprise cooperation motivation AI researchers challenged the nature of the problem, which is precisely the nature of the problem is not a purely academic laboratory environment could ever meet.
Nature of the problem is very naughty, often born in the industrial sector. Therefore, it is difficult under a single conclusion, academia led the industry, academia or industry driven.
According to Linda Hua's point of view, "positive feedback loop" is the best interpretation.
A graduate of Beijing University graduate students told me: "When Dr. MMLab application will be in the list of preferred laboratories, MMLab now known worldwide, comparable to the strength of the Ivy League, even the large class people of God He Kaiming also stayed, so people aspire to. "
Learning and teaching, Q & A, practicing every day, every night to ponder the matter.
"Innovators" and "creator" Linda Hua, in the enthusiasm for scientific research in the dwelling, in affectionate send teaching in plant, he explores the vitality of scientific research, flaunt the spirit of the generals on the battlefield, just waiting for the triumph.
(Finish)
Linda Hua Description:
Linda Hua is the Assistant Professor of Information Engineering, Chinese University of Hong Kong, he was in computer vision, probabilistic inference, and the depth of learning has extensive research experience, and have achieved outstanding results in a variety of subjects. He studied premier meeting and journals published over a hundred papers in CVPR / ICCV / ECCV / NIPS / PAMI such as computer vision and machine. He received the most authoritative international field of machine learning NIPS conference in 2010 Best Student Paper Award, and in 2011 obtained the highest academic conference on Computer Vision ICCV distinguished panelists Award in 2009. He has coached the Chinese University of Hong Kong research team to participate in major international competitions in computer vision ImageNet, ActivityNet, and COCO, access to multiple champion. In addition, he also served as Chairman of the areas of major international conferences CVPR, ECCV, AAAI, ACM Multimedia, etc., as well as top international journals IJCV the editorial board.
Professor Linda Hua in 2012 received his Ph.D. in computer science from MIT in 2007, master's degree in Information Engineering, Chinese University of Hong Kong in 2004, a bachelor's degree in Electrical Engineering and Computer Science, University of Science and Technology of China.
