essay structure
Writing is the level of output, and idea is between input and output. It is your brain to think, innovate, understand deeply, and then generate some new ideas. After you have an idea, you need to verify your idea through experiments, and fully describe your idea through the method part, then you need to use thesis writing at this time.
The importance of writing a thesis
1.1 Papers can carry ideas
Writing a thesis is an essential skill for scientific researchers. One of the functions of the paper is to carry the idea. It is to systematically describe your idea in a predetermined format. Only with the idea can you communicate with your peers.
Why is writing a thesis important? Because when you communicate with other people, it is impossible to describe your idea to everyone every time. If it is presented in the form of a paper, whoever is interested in your research direction can just let him read your paper directly. Your idea has been clearly stated in the thesis.
Therefore, the paper will carry your idea in the most traditional way. It is a communication medium. Others find ideas by reading papers, and you express your ideas by writing papers.
1.2 The thesis is the most direct proof of academic ability
The thesis is actually the best proof of academic ability. If you talk about your academic ability, others may not believe it, but the thesis can be used as the most direct proof. Whether there is a paper or no paper, the treatment and results will vary greatly.
1.3 idea can help the paper to be accepted
In the process of writing, you will encounter many problems, how to express your ideas? How to organize the content? Wait...; Secondly, whether the paper is accepted or rejected has a lot to do with how well the paper is written. The same idea, the same experimental results, presented through different writing logic or organizational methods, may get completely different results—rejected or accepted.
Taking my own thesis as an example, the situation mentioned above appeared in the first thesis I did at that time. When I went to submit some CVs, the reviewers said - I still remember that sentence - large work, but with little novelty , which means "you have done a lot of work, but there is no innovation", and then Turned me down.
But then I went to vote for another top conference. The content, organizational logic, and experimental part of my paper have not changed. The results and methods of the experiment have not changed. The only thing that has changed is how I write this paper. I unfolded various parts around my core idea, did the method part, echoed with the idea, and the experimental part also echoed with the idea. In the end, the same article was accepted, and the reviewers all said that the idea of your paper is very novelty, very new, and very interesting. It's amazing, and this is the impact of "writing", the most direct impact.
It is smooth sailing for no one to submit a paper, but after going through this road, subsequent paper submissions will be much smoother, because with a lot of experience, you will know how to avoid many pitfalls, and you will know what reviewers like to see paper. There is a saying that "three points for idea, seven points for writing", which shows that writing a thesis is really important. For the same idea, one person may be able to get in the top class, while another person may only get in the C group or a fourth district. ,Third District.
2. Why is it not easy to write a thesis?
Writing a thesis is very important, so what is so difficult about it? Everyone wants to write a good paper, but why can't they write well? Why is it so difficult to write, and many times I don’t know how to write?
I summed it up myself, there are a few obstacles:
The first is the language barrier , which is a common problem. Everyone's English writing ability and skills are different, and the papers they write must be different.
The second and more core point is idea conception , you have to let your idea tell a complete story. You didn't have this kind of storytelling ability at the beginning, you didn't have this awareness at all, and then you didn't know how to lower the reader's reading threshold, so that the reviewers read your paper very clearly and very clearly Comfortable, it is easy to get your core idea. If the reviewer can do this, he will have a very high evaluation of your paper.
Third, writing itself is not an easy task , especially for scientific and technological papers. Why do you say that? Because when you think about a problem, it is impossible to think linearly - when you design a method, imagine a tree structure, the root node is your thinking on this problem, this problem has several characteristics, it will Divide into 1, 2, 3... features, and then how to use different ideas or different modules for each feature to solve this problem? In the end you'll have a whole method section sorted out.
But when writing, it is impossible for us to draw a frame diagram or a tree diagram, but we can only write it down logically, from tree-like thinking to linear expression and output. It is difficult, so writing itself is very difficult.
Fourth, we don't have many opportunities to practice . Writing is like swimming, or riding a bicycle. These motor skills have to be practiced. The more you write, at the same time others will give you feedback, and you can improve it, so that you can improve. It is to practice repeatedly in this way, just like shooting a basketball. When you shoot a basketball, every time you shoot a basket, whether the ball goes in or not, it will give you a kind of feedback. Naturally, you will know how to shoot the ball. You can't hit a shot. In the same way, the more papers you write, whether it is the supervisor revising your paper or the reviewer replying to you, it is giving you feedback. Based on this feedback, you can improve and think about the problem. Then solve these problems, and it will be easy to write the paper later.
Therefore, why it is difficult to write a thesis lies in these four aspects. But in fact, the core is the conception of the idea, and the rest are just appearances. The idea of an idea is your writing logic. Academic papers have a writing logic. We go to communicate with our peers. The papers we write need to be logical. It is impossible to write wherever you think of and write a few sentences like those official accounts. The conversation will be broken, or the readers will write as they like.
3. How to make reviewers understand your paper?
Next, I will talk about it from the perspective of top reviewers, how they will evaluate your paper after it is written, this point is actually very important. We can imagine, empathy, know yourself and the enemy, "the other" is the reviewer, your purpose is to convince the reviewer, let the reviewer agree with you through your paper, and then he will accept your article. This is the core point.
Let me talk about the characteristics of reviewers first. Every time a reviewer receives a meeting or a review task for a journal, the meeting will generally require five articles. Some reviewers may have several accounts. Each account After all selections, 10-20 articles may be reviewed at one time. However, there is a cycle for reviewing manuscripts. He has to read so many articles in a short period of time and write review comments. This is a matter of very tight time and a very heavy workload.
At this time, the whole person will inevitably be very busy, and the reviewers themselves have to do experiments, write papers, and submit manuscripts, so the time spent on your paper will be very little, so we must respect the reviewers Make sure to let him get your idea as quickly as possible.
Another point is that the background knowledge and tastes of each reviewer are different.
Nowadays, the number of submissions to AI conferences is very large, and the number of reviewers it needs is also very large, but in fact there are not so many qualified reviewers. Or it is unlikely that your paper will just happen to be reviewed by someone who knows your field very well.
This is an inevitable phenomenon. At this time, the reviewer may not understand your field, but now he is here to judge whether your paper should be accepted or rejected, and to give you a score. Then this At that time, we need to lower the reading threshold of the paper low enough, and introduce the reviewer as a noob, let him get our core idea, and then develop our idea step by step through the paper, he Only then can we get our idea, and then give our paper a good evaluation.
Taking AAAI as an example, each conference will have some established questions. It is hoped that reviewers will score a paper from these aspects. The following are some points that basically every conference will care about.
The first point, novelty, is the attractiveness of the idea of the paper.
Whether it is your concept, the problem you want to solve, or the method you mentioned in the paper, their attractiveness is ranked first. When reviewers attack your novelty, your The thesis is basically useless.
The second point is soundness, rationality.
Whether your paper is technically sound requires experienced reviewers to judge. You can’t make low-level mistakes, you can’t have problems in formula derivation, or some basic knowledge is wrong, or the data set used is not common in this field. Reviewers will use various questions to judge.
The third point is clarity. Is this paper organized well? Or is the paper written clearly? After your idea is formed, you can make a lot of improvements on it, and if done well, it will greatly help your paper to be accepted.
There is also the fourth point, the core claims in your paper, whether your idea is well proved and supported by your experiment , this time depends on your experimental settings, whether it can prove your idea.
These are all around your idea, the influence of your paper, or providing a new data set, or providing some code. These are also considered small contributions, but the core is still the above four points.
What kind of paper can get a good enough review opinion? Let's take a look at a real reviewer's opinion first, and take a CVPR 2022 reviewer's opinion as an example.
This reviewer is very professional, because CVPR is different from AAAI, he did not directly list many questions, let you answer one by one to score, but only one strength and weakness, that is, what are the advantages and disadvantages of this paper what is it
This paper is also my own paper. He wrote these points in terms of advantages, 1.writing, 2.experiments, 3.method, 4.the analysis with MVV, 5.code...It fits perfectly A few points mentioned earlier. This means that the thesis is finished. If a reviewer thinks that your paper has no problems with these points and meets the requirements, then he will give the same review comments as this CVPR 2022 paper.
How can this level be achieved?
Thesis writing process
4. Typical paper structure and writing order
A core paper consists of these modules: Abstract, Introduction, Methods, Experiments and Conclusions .
And my writing order is: method→experiment→introduction→abstract→conclusion . This is the way I think reviewers can get my idea clearly, but it doesn't mean it's the best solution. The reason why I sorted this way and wrote the method first is my reason.
Here is a question: what is the purpose of each module in the thesis? What role does it play for the whole article? When reading, and when you write yourself, you have to consider two aspects, one is from the perspective of "writing" as an author, and the other is from the perspective of "reading" as a reader.
Why did I write in this order, why did I put the method first ? Because I can only design the story line of the idea after I have decided on my writing plan first, so that I can know how to tell the idea, which has nothing to do with how to write the introduction.
The second part is the experiment . The method part is to show how the idea affects the method through a specific plan, and how you formulate the overall plan based on the idea. The second part is that your experimental results will verify your idea. The two parts of method and experiment are also the best to write.
Once you have the methods and experimental sections, you can go back and write the introduction . The introduction should clearly explain the story line of the core idea of the whole paper, what kind of story you want to tell, and how your method and experimental part can confirm or embody the idea.
Reviewers or readers read the introduction first. According to the logic of the introduction, read the story from beginning to end. After eliciting your idea, they will judge whether it is reasonable for you to elicit your idea like this at this time. Is the data of idea new enough? With an understanding of the idea, he will look at your methods and experiments to assess whether the method you designed based on the idea is reasonable enough, and whether your experiments have proved that the idea you proposed at the beginning is effective .
When it comes to the abstract part, I will condense the introduction every time, and then I can write the abstract. The abstract is actually a condensed version of the introduction.
The last is the conclusion . In fact, the conclusion is not that important. It is just a few words, from the introduction of the idea, to how to do the experiment, what is the effect, or some future work, how to do this in the future, or what is your current method? defects, effects, etc.
This is the order in which I write. It is also the writing process of this CAPR paper. Now, taking this paper as an example, I will tell you what his writing process was like and why he wrote it like this.
Sample display and analysis
How should the story line unfold? Taking this CAPR as an example, look directly at its core intro part, which is a bearer of the story line I just mentioned. At that time, I was constantly changing and changing my intro. In the end, the overall logic presented in the intro will be very clear, and reviewers can understand it clearly. The intro part of my entire thesis has two pictures, nine paragraphs, the first picture directly presents my core problem and idea, here I want to solve a pre-adaptation problem, from s to t, also It is a street view taken from a clear street view to a foggy one in another city. The normal situation is to solve it in one step, and the gap will be very large. My core idea is that I divide this large gap into two small ones by introducing an intermediate domain.
I explain why I did this below, because the two pictures of s and m were taken in clear scenes without being affected by fog. The difference between them lies in the style of the two pictures. m and t are the same Shooting in a scene is just because sometimes it is foggy, and sometimes it is not foggy, so the difference between the two domains m to t lies in the fog, so take it as an assumption and think that they There is only one fog gap, which is one of my core ideas, including the problem I want to solve.
At this point, I actually presented my core idea in the fastest and most direct way. After reading my title and looking at my picture, I can know how I want to solve this problem, which makes the reviewer the first I have a relatively clear understanding of my idea for a while, even if I don't know it very well now, but after reading my intro, I must have gotten my idea very clearly. And after reading the intro, I will know why I brought up the idea like this, and I will use other experimental data to prove that my idea is reasonable. At this time, the reviewers will basically have a very accurate grasp of the paper. If I The methods and experimental parts are well done, then basically my thesis will be awarded.
When the reviewer has a positive impression of you, the review comments he will give you will definitely be positive.
As an author, how did I write this paragraph? Then it must be clear, what is the purpose of each of our paragraphs? What kind of logic should be written? The first paragraph of the introduction-p1 here explains my purpose and proposes the problem I want to solve, but it is not enough. I will continue to emphasize my problem at the beginning of the intro part. Because for the reviewer, he must figure out what I am doing and what problem I am trying to solve. If he doesn't even understand this, how can he understand my article?
So I was emphasizing my problem from the beginning-semantic segmentation in foggy weather is actually very important for autonomous driving. This problem can be solved very well in clear scenes, but there will be many problems in foggy scenes, because the fog will cause the visibility of the pictures taken by the camera to be much lower, which in turn will cause various problems such as poor performance. Add to that the fact that you have a lot of data and artificial labels in clear scenes, but not in harsh foggy ones, which makes the problem even more troublesome. And what I want to solve is the semantic segmentation problem of SFSU foggy scene.
Based on these two points, I need to transfer some knowledge of scene segmentation with clear normal weather to unlabeled foggy scene pictures, so what I want to solve is this kind of transformation. I completely cited this problem in the first paragraph Come out, this is also the most important role of the first paragraph - let the reviewer get exactly what problem I want to solve, and then he will think about it with this problem, he will think, um, it's interesting, it's really hard to solve .
I have emphasized this difficulty before. The difficulty is that the low visibility of fog will cause more trouble than clear pictures. At the same time, there is no data, and the cost of labeling is very high. It is impossible to use a large amount of data to label. At this time, a common idea is to transfer knowledge. How to migrate?
Look at the second paragraph again. At this time, we will introduce the first type of problems, the first type of solutions, and what problems exist in the corresponding solutions.
Here I talk about Intuitively, that is, intuitively, what I can think of immediately, the simplest solution to this problem, domain adaptation, directly uses the traditional pre-migration solution to combine clear image and fog image, that is, normal weather It is the most intuitive idea to directly transfer the gap to solve the problem of the pictures and the pictures of foggy days with a traditional pre-adaptation method. At this time, I am guiding the reviewers to follow my train of thought.
The problem with the first type of solution here is that it uses a kind of confrontational thinking, including self-supervision thinking, and there will be a shortcoming, that is, they ignore how the gap is formed and what the gap is caused by. Yes, regardless of whether the gap is a fog gap, a rain gap, a style gap, or a variety of weather gaps, it treats it as a gap and handles it uniformly instead of solving it in a targeted manner. I also mentioned below that other papers have proved that it is not good to directly use domain adaptation to solve the problem of foggy weather, because the gap in foggy weather is relatively large, and it is not that simple.
The reviewer will think after reading this, yes, it is really not good to solve it directly with the DA solution. It is true that he has not considered how the gap in the fog is caused, and he will want to continue reading.
In the next third paragraph, I immediately changed the painting style and began to discuss how to solve the problem, so that everyone can focus on the fog, because the segmentation of the foggy scene must be directly related to the fog, and everyone will think that This is the most direct impact. If the problem of fog gap is solved, this problem can be solved.
There are also many people who have done various researches and proposed two types of solutions. The first type of solution is, can I remove the fog in the foggy pictures? I used some fog removal solutions to remove the fog. After removing it, wouldn't the effect disappear? However, the ideal is very full and the reality is very skinny. The reality is that these defog solutions can’t do this well at all. While they can remove part of the fog, they will also lead to various problems. The things are also removed, and the result of semantic segmentation will not be good.
The second type of solution is to add fog. Add some synthesis to the clear picture scene. After adding synthetic fog, the clear scene is labeled. Can we do a supervised training? But the problem is that the added synthetic fog is different from the real fog, and there is also a gap between the synthetic fog and the real fog, which will lead to a bad final effect. So I emphasized at the end that we think that these methods are all too much emphasis on the fog factor, while ignoring one other factor, and that factor may be very helpful to solve the current problem.
So which point is it? We continue to reinforce the difference. After reading the above paragraph at this time, the reader will realize that what other factors may affect this problem? He will follow your train of thought and continue to think about it, thinking while reading.
At this time, we should start preparing to lead out our idea. This lead is also a successor to the problem of the solution in the previous paragraph. Let’s think about it from another angle and propose that we should investigate the domain gap in an intuitive way. This is to avoid To deal with the entire gap directly, and at the same time avoid using these synthetic substances. So what do I think it's like? I think it is caused by two factors. One is mixed, which has both the effect of fog and the change of style. It is the combination of two factors that makes this problem difficult to solve, and these two factors are very important to this problem.
Later, I emphasized it again, and then expanded the point of view just now, that is, I think it should have two gaps, and we can separate these two gaps and solve them separately, through an intermediate domain, this is Our core idea.
So far, the reviewers must know very clearly what kind of idea I used to solve the problem. The key is that I will use experiments and discoveries to prove to the reviewers that they can indeed be separated. A large gap can be divided into two small gaps, one fog gap and one style gap. Why can they be separated? I will tell you through an experimental and intuitive phenomenon, which is equivalent to confirming the rationality of my idea.
Then go to the next paragraph. At this time, I used the second picture in my thesis. This picture gave a clear explanation of the most important motivation of my idea.
The core idea is that we have found a kind of mvv index to measure the gap between different domains through a numerical quantitative experimental analysis. At this time, I found that for the dual gap, if I close it alone, it will also It is to reduce the style gap. For example, after I pass the data of m from s, after it learns, I can reduce the style gap from 0.089 to 0.067, which is the part of the yellow line in the picture, but at this time my fog gap is actually not How to change, still from 0.033 to 0.37. What is the problem at this time? I can actually solve the style gap alone without affecting the fog gap. This is a very useful thing. Previously, a large gap was not easy to solve. I divided it into two small gaps and solved them separately. I verified through experiments that they can be separated, so this is very reasonable.
After reading this, the reviewer will feel that my idea is very reasonable, and then he wants to see how this idea works? The recognition of my idea will affect all his subsequent evaluations.
In the fifth paragraph, I wrote a special chapter called motivation, which explained why I came up with such an idea, and how did I come up with the idea? The findings in the picture just now proved the rationality of my idea, including the fifth paragraph, the sixth paragraph, and the seventh paragraph. I used three paragraphs combined with that picture to show the reviewer the validity of my idea. rationality.
Then based on the idea and a discovery of the experimental phenomenon just now, I will start to refine my idea and use a paragraph to point out the design principle of my method. The logic of this paragraph is very clear, step by step to tell you what kind of design method my idea uses, and finally how to solve the problem I want to solve step by step.
The last paragraph is to summarize the contribution of your thesis. The summary contribution is generally three points, what methods were used before, and then what new discoveries have you made, and then whether your methods are very effective.
In the third point, I said that our method is far beyond the current search results. On these three data, I also expressed that my method has a set of generalization ideas. My idea can not only be used in In foggy days, it can also be used in rainy and snowy days, which is another point. These three points are my core contribution.
Looking back now, every paragraph I said in the intro part, in fact, the most important thing is the logic of storytelling. To sum up, the first step is to draw out the problem; the second step is to introduce the existing solution to this problem, and then point out There are some problems in this solution; in the third step, you can propose your own solution, that is, your idea, and introduce your solution in detail, and then verify your idea through a series of experimental data, so that reviewers I feel that your idea is not based on brainstorming, but has been analyzed and verified by experiments, so it is relatively reliable; the fourth step, after explaining the process clearly, you can continue to refine and explain your method based on the idea. Point out the core principles of the method, and then refine the design of the method. This method will make the logic of your entire paper very clear, and the reviewers will feel very reasonable after reading this. Based on the verification process of the previous idea, the reviewer will feel that it is very reasonable. The new discovery and how your method part will eventually improve the effect will also make him feel that your idea is very novel.
And the reviewer's review comments are basically written following your thoughts.
Now when we look at the writing of the thesis, there are actually a lot of details. I only talked about the intro part. In fact, whether it is your method part or the experimental part, there will be many, many problems, and you have to find a way. Grab the reviewer's attention, set up some obstacles, and then tell him how you solved the obstacles, and tell your reviewers what my idea is like through detailed experimental settings and experimental results. People will judge for themselves.