Research Matters (Ten Thousand Words Long Essay, Sincerely Shared) Unauthorized reproduction in any form on websites, forums, or communities is strictly prohibited without the permission of “AI Technology Review”! For official account reproductions, please first leave a message in the “AI Technology Review” background to obtain permission, and when reproducing, please remember to indicate the source and insert the business card of this official account.
Author: Han Xiaoguang
The corresponding English word for "科研" is "research". The "re" in it means "again" or "repeatedly", and "search" means "to find", so "research" can be understood as "finding things repeatedly." I heard this during a lecture by Academician Shen Xiangyang when I was a student, and it left a deep impression on me. So, what exactly are researchers trying to find? Usually, they are looking for two things: problems and methods. It is easier to understand the search for methods, as the general perception of research is that it is about finding ways to solve difficult problems, such as searching for methods to treat cancer.
So why do we need to find problems in scientific research? To answer this question, we first need to understand what a "problem" is. Based on years of research experience, I believe that "problems" generally fall into two categories: one is problems that the academic community is widely researching, such as the controllability issues in video generation currently; the other is problems that have not been extensively studied yet, which are in the nascent stage. In this case, it is generally necessary to find problems through demand-driven approaches (for example, the demand for floor cleaning has driven research on robotic vacuum cleaners), logical reasoning (such as extrapolating from how to generate images to how to generate videos), and thorough exploration (such as continuously contemplating what neural network structure is the best). For these less obvious problems, when we unearth them from complex contexts, further refinement and elaboration are needed (that is, defining the problem thoroughly and rigorously, for example, using mathematical language to describe the input and output of the problem).
If we treat scientific research as a profession, the above is an explanation of the job contents for this profession. Based on my experience over the years, I prefer to understand scientific research as a mindset, a mentality towards facing challenges. For example, when I used to play the game Counter-Strike in my early years and experienced dizziness, my choice back then was to give up on the game. However, I believe there is another mindset which is to first understand why I was feeling dizzy. Is it a personal issue or do many people experience dizziness? If it is a common issue, then research how to improve the game to solve this problem. Therefore, scientific research is actually everywhere.
Of course, from a personal perspective, if you delve into everything, it may feel a bit like sweating the small stuff. We can choose to overcome more appealing challenges instead. As long as we keep embracing new challenges, we will continue to learn new things!
This question may be the most frequently asked question by students I have encountered in my teaching career. When I first started as a teacher, I didn't know how to answer this question. Because every time I think about my past, I rarely had this kind of doubt. From obtaining a master's degree, working as a research assistant, pursuing a Ph.D., to finding a teaching position, for a long period of time, whenever faced with choices, I always unhesitatingly chose to continue research because there were always interesting topics left unfinished. After many years of research experience, trying to answer this question, I still find it difficult. Because in order to know whether you are suitable for research, you not only need to understand research, but also need to understand yourself. The latter may be more difficult than the former.
To provide some reference, I would like to share some differences between the process of conducting research and the typical learning process. Here, the typical learning refers to college course studies, completing assignments, group projects, or participating in competitions, etc. Generally, these types of learning have specific goals and answers, so when you encounter difficulties in the process, you will clearly know that there is a problem with your implementation process, and you can proceed to pinpoint the problem. In research, there are many factors that can lead to difficulties, such as problems in your implementation, poor design of the solution, or setting incorrect goals from the beginning, making it very difficult to pinpoint the problem.
And when you keep searching without finding any problems, self-doubt will arise, doubting your abilities, questioning the rationality of the topic, wondering if you have chosen the wrong direction, questioning if you should not have pursued a PhD, and so forth. At this time, it requires strong conviction and the ability to persuade oneself in order to continue moving forward. Most of the time in scientific research is not smooth, it is a process that involves continuous self-doubt, and also requires constant self-persuasion. From this perspective, scientific research may be more suitable for optimistic people with strong beliefs.
In the eyes of everyone, the research process requires a lot of abilities, such as the ability to generate ideas, write code, write papers, give presentations, and so on. But this doesn't mean that only "hexagon warriors" are suitable for research, because research is a collaborative effort, and you can excel in a few of these abilities and still do well. Therefore, there is a common misconception among undergraduate students: if my grades are good, I should naturally continue to pursue a Ph.D.; if my grades are not so good, then I'm not suited for research. The best way to judge whether you are suitable for research is to give it a try. You can participate in a research group during your undergraduate studies, pursue a master's degree after graduation, or work as a research assistant for a period of time before pursuing a Ph.D.
I think there are two criteria that can serve as references when you make a choice: the first is interest - if you are interested, then go for it; if you are unsure if you truly like it, as long as you are not against seeing if the following matter presents challenges to you, if there are challenges, then you might as well give it a try. With challenges, you will definitely gain something in return.
Many people often think of publishing papers when it comes to scientific research, thinking that the purpose of research is to publish papers. I used to think the same in my early years. This may be because papers are often used as evaluation criteria, and when it's difficult to judge the quality of a paper, quantity becomes the most intuitive indicator. Therefore, a research paradigm for the sake of papers has emerged: first, read a bunch of papers, then summarize the patterns in the papers and form a template, and finally, apply the template, treating research as a fill-in-the-blank exercise. I think templates are a great thing because they provide certain standards and guidance. For example, after reading a large number of papers, you will summarize in the "Introduction" section: What problem you need to address? Why solving this problem is important for the entire field? And what is the contribution of your work to this problem? Attempting to answer these questions will indeed be of great help to your own research.
In fact, what I don't like is treating a thesis as the goal of scientific research. I believe the goal of scientific research is only one: to explore the unknown and generate knowledge! And once you have generated knowledge, you need to present it in a standardized form, namely, a thesis. Therefore, a thesis is just a summary and intermediate product of the long scientific research process. Of course, a thesis also serves a very important function, which is to publish your views for communication with the community and peers. I like translating "论文" as "paper". Even in the field of science and engineering, a thesis needs to have arguments supported by evidence. For the natural sciences, the evidence may be mathematical derivations, and for engineering, the evidence may be experimental verification. I remember when I first started research, I used to enjoy reading the methods section of papers very much. When I encountered a good article, I would admire the clever design and elegant theory in it; but once I became very familiar with the methods in a particular research field, I would prefer reading the experimental and results sections of papers, often unintentionally seeking out flaws, hoping to find problems and my own research direction. And at a certain stage, I started to enjoy reading the "Introduction" and "Related Work" sections of papers more, as authors often express their attitudes and viewpoints toward the entire research field there. Nowadays, with the explosive growth of articles in the field of artificial intelligence, fewer and fewer scholars are dedicating themselves to writing papers, and it is rare to see exciting viewpoints in papers.
If we consider a thesis as a means of communication, there is no need to be too strict about its form. I have heard that many scholars pay great attention to the use of words and even to the extent of punctuation marks when writing. Of course, I truly respect this craftsmanship spirit, perhaps inheriting from the older generation of scholars, because theses in the past represented human civilization and needed to be stored in libraries. However, in the current electronic age, I personally do not think it is still necessary. Just like how two people speaking different dialects can communicate well. Similarly, for communication, a thesis only needs to clearly convey the viewpoint and provide sufficient evidence for the viewpoint, and that is sufficient.
During the process of forming each paper, whether it is defining the problem or proposing a technical solution, authors often need to go through multiple attempts and iterations before arriving at the final version. The papers that we see usually only show the final form. For many researchers, especially students who are just starting out in research, the multiple failed attempts behind these papers are the real valuable wealth. This is also the original intention behind my initiation and organization of the "Stories behind the Paper" lecture series. Currently, the lectures take the form of interviews with well-known scholars sharing their insights and experiences during the paper-writing process. In fact, I have been constantly thinking about how to present more of the attempts and stories behind papers in various forms.
In the summer vacation of 2008, I obtained the qualifications for recommended admission to graduate school and confirmed that Dr. Liu Ligang from Zhejiang University at the time would be my master's supervisor (Dr. Liu was my mentor in scientific research and later moved to the University of Science and Technology of China). Under Dr. Liu's guidance, I started to delve into the field of scientific research. During the initial period, Dr. Liu assigned me two tasks: one was to write code to implement classic computer graphics papers (such as Poisson Image Editing); the other was to read papers related to Mesh Segmentation. Whenever I think back to that time, I am filled with a sense of excitement. When discussing this, two moments come to mind: one night during my senior year, I was struggling to solve a stubborn bug. Everyone else in the dormitory was asleep, but I couldn't fall asleep. After tossing and turning for two hours, I suddenly thought of a possible issue, so I quietly turned on the computer and quickly resolved it. The thrill of that moment kept me awake all night. Another time, I was hanging out with friends when my mind was full of papers on shape segmentation. Suddenly, I saw a sculpture of an elephant and started thinking about how to segment its trunk. As I pondered this, I realized that I may have misunderstood a part of a paper I had read before, leading to a logical inconsistency and blocking my train of thought. At that time, I hadn't learned how to shift my focus, so I was constantly troubled by the problem and couldn't enjoy the outing. Eventually, I found a nearby internet cafe, downloaded the paper, studied it for a long time, and finally figured out the issue. Looking back now, the thirst for knowledge at that time, the excitement after solving a problem or learning a new piece of knowledge, all felt so pure.
This kind of purity has always been maintained throughout my postgraduate and doctoral studies. In 2018, I started my teaching career and began conducting research with students, with one of the most important steps being choosing what kind of topic to work on. Initially, I would be driven by curiosity to identify research questions that I found meaningful based on papers I read or presentations I listened to from others. As a result, my students and I gradually started to publish some good results in top conferences in the field of computer vision, such as CVPR, and occasionally we were selected for Oral Presentations, and twice for Best Paper Finalist. When I noticed that oral presentations and best paper nominees received more attention, I started researching the commonalities among these papers and tried to summarize what kinds of papers were selected. I then used these patterns to guide us in selecting future topics. However, it seems there is a curse in everything - the more you pursue something, the more elusive it becomes. I have also tried to summarize what kinds of papers go viral (i.e., receive widespread attention on social media), what kinds of papers have high citations, what kinds of work being open-sourced will garner more GitHub Stars, and what kinds of work are more likely to be implemented and bring support from the industry. Since then, curiosity seems to have become less pure: when encountering an interesting problem, I unconsciously ask myself, can working on this lead to the best paper? How much attention can it attract? What are the chances of practical applications?
So how can we return to purity? This question has always troubled me, it's really hard. If I had to give some advice, I think we need to work on blocking out the surrounding noise: who published a few more papers, whose work has brought breakthroughs (reported by the media), who has won an award, and so on. Then focus our energy on the problem you are currently solving, thinking about how to solve it well. Also, carve out some time to read papers, to find that initial state of learning new knowledge.
Since becoming a faculty member, I have had communication with many doctoral students, including those who are not my own students. The most common question they ask is probably: After graduation, how should I choose between academia and industry? On this issue, I have heard two different opinions: some people feel that the academic papers published are worthless, and most of the research done has no value, purely for self-entertainment; while others believe that in the industry, all tasks are about solving engineering problems, with no room for innovation.
In the past few years, although I have been in academia, I have also collaborated with many industries and have students involved in entrepreneurship. To some extent, I have formed my own understanding. I believe there are three different roles related to this issue: researchers, entrepreneurs, and research and development personnel in large companies. Of course, in many large companies, there are indeed researchers. In terms of goals, researchers generate new knowledge through scientific exploration; research and development personnel in large companies need to use knowledge to create more wealth; while entrepreneurs explore how to turn new knowledge into wealth.
From the mission bestowed by society's perspective, I believe that the academic community only needs to consider how to create new knowledge.
Understanding while not needing to be responsible for whether this knowledge is useful or can turn into money. In fact, as long as there is curiosity, continually raising questions and exploring to find the answers will lead to knowledge. Therefore, in the academic world, curiosity and the ability to freely explore are essential. Mentally, one must also have the ability to endure waiting and the ability to self-motivate. The mission of the industry is to explore how to transform knowledge into money, which may also lead to the creation of new knowledge in the process. From a societal perspective, I think the industry also needs to allocate some money to support the academic world to create more knowledge, thus forming a positive cycle.
After understanding the above, in choosing between academia and industry for a doctoral student, personally, I think: if you want to become a researcher, you need to have the ability to self-motivate and the ability to maintain curiosity. There are two types of researchers, one in universities and the other in research institutes or corporate research departments. If you aim for a position in academia, in addition to what was mentioned above, you also need to have patience, that is, the ability to cultivate patience in teaching students. In addition to creating knowledge, a very important mission of academia is to educate students. However, there are new students every year, which means you may have to do similar work year after year. Therefore, it is difficult to handle without a mindset that does not become annoyed. If you want to enter the industry, you need execution capability. This capability is not just the ability to complete tasks, but it is also about your acceptance and even willingness to do tasks that you may not agree with or find meaningless. I believe entrepreneurs need more comprehensive and diverse abilities. One very important ability is perseverance. You need to have strong resilience to "hang in there" in the face of constant failures and glimmers of hope.
The process of scientific research is full of uncertainties, and these uncertainties often make me suddenly feel a chill down my spine. For example, during the execution of a project, a sudden problem arises that I can't figure out on my own, which may cause a project that has been ongoing for a long time to suddenly lose its meaning to continue. In fact, the entire process of scientific research is a continuous process of asking oneself questions and answering them through logical reasoning or experimentation.
In a specific project execution, we need to learn to ask ourselves questions, such as: Why are you doing this project? What are your expected goals? What kind of contribution do you hope to make to the community? What are the possible technical paths to solving this problem? What are the possible technological innovations? And so on. As for how to ask yourself questions, I think one principle is to always ask yourself why. Approach it with a critical mindset, constantly challenging yourself (I also often challenge students). Only by being tough on yourself during the process will reviewers be friendly to you. In addition to learning to ask yourself questions, you also need to make an effort to answer your own questions. You can conduct research with questions in mind, design experiments to validate them, discuss with peers, seek help from mentors or seniors, and so on.
When asking yourself questions, in my past research experience, I found two questions that are very easy for students to overlook and very difficult to answer. The first one is, when you come up with a new question or a new approach to an old question, you may want to try asking yourself: why has no one done this before? Is it because others really have not thought of it, or do they think it is meaningless, or have they tried and failed? In my past experiences, I have often found that what you think is new, upon further investigation, you may discover that there are already many related works, and even some very early works that have attempted the same thing. And when you cannot find related works, it is indeed possible that someone has already tried and failed, making it difficult to answer this question. However, trying to think about the reasons and logic behind "why you could come up with it, and why you are able to do it in the end" may help you better understand the significance of doing this task.
Another frequently overlooked issue is that when a technical solution has been determined and proven to indeed improve effectiveness, many students may think that it is sufficient and start considering writing their paper. However, at this point, I feel that we still need to ask ourselves one question: Is this already the best technical solution? Are there no other possible better technical solutions? Answering this question is not easy because you need to consider all possible technical solutions. Of course, during the project execution, we may not need to come up with all possible technical solutions. But thinking about this question will definitely deepen your understanding of problem-solving and is likely to promote further improvements to your technical solution.
In the actual process of scientific research, we may ask many questions, but these questions do not need to be overly detailed. Otherwise, we would be getting into unnecessary details. In fact, we only need to choose some important questions and make our best efforts to try to find the answers to these questions. From past experiences, we have found that when completing each paper, there are actually many questions for which we do not know the answers, but this does not affect the draft of the paper. And those remaining questions are precisely the starting points for further work.
In the process of collaborating with students in the past, I found that there are two different ways of working and thinking among them: the first type, when solving a problem and designing an algorithm, they must first clarify all the technical details before starting to write code to implement it; the other type is to dive right in without much planning, start coding directly, and then troubleshoot as issues arise. I will refer to the first type as mathematical thinking and the second type as engineering thinking. In fact, both approaches can solve problems.
When working on a complete project, these two ways of thinking will also lead to significant differences: the first one will spend a lot of time thinking about problem definitions, algorithm design, how to be innovative, and so on. Since this kind of thinking often takes a lot of time, not starting the implementation promptly will result in very slow project progress; the second one will start executing without much thought, solve problems as they arise, and deal with issues as they come up. Although this approach speeds up project progress, in many cases, due to the lack of contemplation, the work itself may become routine and lack innovation.
In this day and age, there is a huge surge of articles in the field of artificial intelligence. I have communicated with many students and found that the majority of them approach their papers purely from an engineering perspective: they choose a topic, then look for the latest papers from top conferences on that topic, find the code, try to run the code, identify algorithmic problems, attempt various methods to solve them, and only start writing the paper once they have improved the results to a certain extent. Personally, I really dislike this approach because the work produced in this manner often lacks scientific value. What worries me even more is that if students submit papers in this way and get accepted (in fact, many such papers are accepted these days), they may feel that publishing articles is easy, which could diminish their respect for scientific research.
However, if this approach is only seen as a way to exercise engineering abilities for younger students, it can also be a good method. I have indeed encountered students who have developed strong engineering skills through this approach. They seem capable of achieving results as long as they have a reasonable plan. However, they also have a significant issue: when facing a problem during a project, they immediately focus on overcoming it and may become fixated on it, even if it is not a critical issue in the entire project.
Many people have some degree of confusion about what their interests are or what types of work are suitable for them. In fact, there is a criterion: whether you feel a sense of achievement when doing something. For example, I don't like to play video games because I die quickly every time, playing games doesn't give me a sense of achievement.
In scientific research, many people enjoy conducting research because solving a problem can bring them a great sense of achievement. However, in research, there are more failures than success. Even a strong researcher like Howe Kah Ming once said that 95% of the time spent in research is a failure, and only 5% is for enjoying the sense of achievement. If we define the solving of a problem and the publication of a paper as success, then many students actually spend a long time without feeling a sense of achievement, especially lower-grade students. Among my own students, there are many cases of papers being rejected repeatedly or not having any papers published within two to three years, which was also my own experience as a student. Being in a state where one cannot achieve a sense of achievement for a long time can be very challenging, and many students actually give up pursuing a Ph.D. because of this tremendous pressure, leading to anxiety and even depression.
I have encountered several of them and really want to help. However, this kind of mindset issue requires individual adjustment. I think the essential problem here is how each person defines success and a sense of achievement. If you believe that only by completing a thesis can you achieve success and feel accomplished, then you will likely lack positive feedback for years. Therefore, we need to break down the sense of achievement. We can divide a project into several modules or a few questions. When you solve a problem, it is a success, and you may receive positive feedback within a few months. If you solve a bug or understand a paper, even if it is just one success, you may feel a sense of achievement every day. If you don't consider the thesis itself as your research goal but treat learning as the goal, as long as you are more knowledgeable today than yesterday, you will feel a sense of achievement.
Of course, as you learn more and more, if you don't make some changes, the sense of achievement you feel will gradually weaken. At this point, it indicates that you may need bigger and more challenges. You can start trying more difficult problems, or start exploring a different direction, and so on.
The viewpoints mentioned above, I have brought up with many students before. Some students have given feedback saying, "This is very 'chicken soup'!" But I personally believe that we do need "chicken soup" because it can indeed provide us with some energy.
Everyone knows how to say "to guide" students, in the early days my understanding of guiding students was this: mentors have some ideas and research goals they want to achieve, and they need to guide some students to help them accomplish these goals together. In the early days of my teaching career, this was indeed how I carried it out. Almost all the projects the students worked on were my own ideas, problems I wanted to explore. For them, they did gain experience and were trained during the process. But now, looking back, I realize that I treated the students as tools, letting them help me achieve my dreams.
During that time, I also encountered many students who had their own ideas and goals that they were interested in achieving. There was a student whom I communicated with for a long time, advising him to execute according to my thoughts, even though he agreed. However, after the communication, he still carried out the task according to his own ideas, which was indeed frustrating.
Regarding how to guide students, I have asked many people. I once asked Professor Liu Ziwei from Nanyang Technological University (he is a student of Professor Tang Xiaoou from The Chinese University of Hong Kong) how his advisor guided students. I received one word, "tolerance." I also asked Dr. Li Zhengqi, the first author of the best paper at CVPR 2024, how his advisor (Professor Noah Snavely from Cornell University) guided students. One thing that left a deep impression on me was that his advisor would.
There are some topics and ideas that I am particularly interested in and have been trying to sell them to my students to work on, but I have not been successful.
Based on the above experiences, it made me rethink the relationship with students. As a researcher, we need to guide students in our team to work on topics that I personally find meaningful. As educators, we need to support students more in achieving their own success. Regarding the relationship between mentors and students, I don't really like the term "guidance," I prefer to work with students in a cooperative way to achieve success together, learn together, and grow together.
As a community grows larger and resources become limited, rules for resource allocation are needed, leading to the establishment of an evaluation system. The academic world is no exception. With an evaluation system in place, scientific research becomes purposeful. Having guiding objectives is a positive thing, but we need a good evaluation system. When the evaluation system is flawed, guided objectives may lead to cutthroat competition. At such times, many scholars are forced to do things they are unwilling to do in order to survive. This issue has troubled me for a long time. However, at some point, I inadvertently heard a phrase: "Scholars still need a bit of integrity." This sentence instantly comforted me a lot. Scholars today were literati in ancient times. I strongly agree that "scholars also need a bit of integrity," in simple terms, "do not lose yourself."
Summary: Lastly, I would like to share two things discussed in the previous group meeting with everyone again: First, conducting scientific research requires having "high aspirations, quick hands, and a broad mind". Having high aspirations requires us to think more, set high goals; being quick-handed requires us to be diligent and strengthen our execution ability; having a broad mind requires us to be a bit rough in dealing with short-term failures and not overthink. Second, frequently ask yourself: "Do I still have a strong curiosity about unknown things?" "Do I still feel excited when solving a problem, even a small one?" When you notice your curiosity and excitement diminishing, make efforts to protect them!