As a scientific researcher, what you must do every day is to read papers and input them. But in the face of massive literature, how to input efficiently without missing information? This process requires us to master some literature reading skills.
It is often said that literature is intensively read, but how to do intensive reading? How to understand the development logic of a paper? How exactly does the author express his core work?
1. Why read a thesis
☞The Importance of Reading Papers
①Absorb the author's idea: find new content in the paper
In fact, the process of reading a paper is the process of absorbing the author's ideas. We must first know what the existing research is like, and then look at what new content is proposed in the paper. Only after we have these can we think of our own ideas Do a good job of paving the way, this is one of the sources of our idea,
② Quickly understand a field: through related work and references, clarify the direction of follow-up learning
Usually we learn about the latest status of this field through the latest papers in the research direction and related to the research direction; usually the paper will summarize the current research status under this issue and list all the related work. Including what methods are used in each work, etc., through these can help us to clarify the latest status of our research direction.
So, how to get more literature? For example, in the process of reading a paper, we will suddenly discover a point of interest. It turns out that a certain problem can be solved by a method we did not expect, as if a new window suddenly opened, then we can "follow the vine" Find more related literature to read. We can go to the references part of the paper, then to the references of the references, and then the references of the references of the references... keep nesting dolls, so that we can do a backtracking in this way. Through this method, the background knowledge and development context of a direction can be sorted out clearly. This process may be very time-consuming, but it is necessary, because you will learn a lot of new knowledge, whether it is background knowledge or theoretical knowledge, and you can quickly understand a field.
Of course, if you want to quickly understand a field, reading the literature review is the fastest way. The literature review will introduce related methods in more detail, including the data sets we care about.
③Study experiment settings: data set, evaluation index, SOTA performance, ablation experiment, visualization
In the process of reading the paper, you can also learn how the author designs the experiment; when the author solves this problem, what data sets are used, what evaluation indicators are set up, including existing research questions, and the performance of the SOTA solution provided by the author Where is the step, and how did the author design the ablation experiment—how did the author verify it through experiments step by step, how each process in the method affects the final result, these will be presented in the experimental part .
Another point is some visual analysis, because charts are clearer, more intuitive, and easier to catch people's attention than text, so we can also learn how others present their work through papers.
④Learning writing logic: how to develop the introduction, how to classify related work, how to highlight key points, and how to guide readers
Learning the writing logic of papers is also a very important point in reading papers. When we look at the papers that are most published in top conferences or top journals, the version we can see is actually obtained by the author after many revisions and iterations. At the same time, it may After many rounds of revision and improvement based on reviewers' opinions, you can learn a lot from a relatively perfect version of the paper after so many improvements.
For example, how does the author explain in the introduction part, how can he quickly catch the reader's attention; how does the related work part classify; A series of collaborative methods such as advanced typesetting and charts to guide readers... These can actually be learned. So reading a paper, you can learn a lot from it.
☞Why is it so difficult to understand the paper?
①Language barrier: Insufficient English reading ability
Everyone's English proficiency is different. Some people can start reading English papers quickly, some people may need to use translation software, and some people may still have a half-knowledge after using translation software.
②Background knowledge: lack of basic background knowledge and theoretical level
Especially when beginners first start reading papers, it may be very painful, because of the lack of accumulation of background knowledge, the author writes a lot of professional vocabulary and terms, you may not know what they mean at all, and sometimes you need to understand the papers. Some theoretical knowledge, because the author defaults that you have mastered these theories, or due to the limitation of the length of the article, it is impossible to write all the theories, only some related to this work, but you will know when you read Lacks a lot of background support. It will take a lot of time to read a paper, and the effect is not necessarily good.
③Reading method: Intensive and extensive reading methods are not clear, unable to grasp the key points
How to read intensively/extensively? How to capture key information? There is logic and method to these. If you don't master these methods, you won't be able to grasp the key points when reading the paper.
④ Quantitative change to qualitative change: reading a lot of papers is time-consuming, and you need to persist and develop habits
We often say that we need to read a lot of literature, but how "a lot" is this large amount? This actually varies from person to person.
Maybe many people can only say that they read extensively, and they don’t read many papers thoroughly. Of course, you don’t need to read every paper well, but the papers that are very important in your research direction need to be read. Spend a lot of time reading through, because it will directly affect the quality of the ideas you can think of. In the process of reading through the paper, you will have some understanding of your own. This kind of reading is your deep thinking and deep thinking. understand.
The quantitative change to the qualitative change here is when everyone reads a lot of papers, and then gradually achieves a deep understanding of the papers in the process.
⑤Communication and discussion: Lack of timely communication and guidance when encountering unclear points in the paper
When you encounter unclear places when you are reading a paper, you may not be able to communicate with others in time to resolve doubts; or for the same problem, everyone may have their own opinions and ideas, if you are only limited by yourself If you don't listen to other people's understanding, then you can't broaden your thinking field.
We will definitely encounter many problems in the process of reading papers. If you happen to have senior brothers or sisters who are much more experienced than you, you can communicate with them and ask for advice; they may give you some guidance and make it easy for you. Read the papers, and the brothers and sisters may have a deeper understanding of certain issues. When communicating, you will get some inspiration and gains.
Thesis intensive reading tips——
When we read a paper, we should read the paper with questions. Only in this way can we supplement the foundation and understand the scheme or method from reading the paper. We can begin our reading with the following seven questions:
- What problem does this paper address? What scientific hypothesis do you want to test?
- What research is relevant to this paper, and how are these studies classified? Which researchers deserve attention?
- What is the solution mentioned in the paper? Where is the key point?
- How are the experiments in the paper designed? What conclusions were drawn from each experiment?
- What is the dataset used for quantitative evaluation? (If the code is open source, find the link)
- What exactly is the contribution of this paper? (Explain within three sentences) Where is the new?
- What else can be done based on it next? Is there any work that can be continued in depth?
Through these seven questions, we can evaluate the level of understanding of the paper in the process of reading the paper. If you can answer every question clearly when reading a thesis, you have already understood a thesis quite deeply.
2. Typical paper structure
For a research paper, it will mainly be divided into the following parts:
·title
Include this article to research the problem and its core idea.
In the title part of the paper, be sure to use the most concise language to highlight the most important points of the paper.
·author
The name of each author, in order of first author, second author...
·Affiliation
Author's affiliation
·Summary
The abstract is the condensed essence of an article, so every sentence is useful. For example, let's take a look at the abstract section of this paper.
Taking a CVPR paper as an example, let's take a look at how he wrote it.
Firstly, explain the difficulty of the problem to be solved, and from what angle we can solve it based on this analysis. At the same time, briefly mention the existing solutions, and then naturally lead to the core idea of this paper, and finally expand the core idea, highlighting the advantages and contributions.
In this process, pay attention to which linking words the author uses to make a transition and progress between each layer of meaning...
Whether we are doing intensive or extensive reading of the paper, the abstract part is what we need to focus on, because from the abstract, we can see the core focus of this work. Sort out what contributions it has made on the basis of its predecessors.
·introduction
Also use this article as an example. Let's take a look at how the author develops the story line in the introduction.
The author writes according to the structure of the total score.
The author first uses illustrations to illustrate the main idea of this work, and achieves the effect by adding an intermediate domain m.
Then the first paragraph draws out the problem and analyzes the angle of solving the problem: SFSU→DA.
The second paragraph describes the first type of solution to the problem and the remaining problems in it: DA×.
The third paragraph describes the second solution to the problem and its shortcomings: Fog (defog/add fog) ×.
The fourth paragraph takes over the problems in the previous solution and puts forward your own idea: Fog+ Style √
The fifth to seventh paragraphs use an experimental discovery to prove the rationality of their idea: Fog+ Style √
Eighth, refine your own idea and point out the design principle of the method: disentanglement+relationship→SFSU.
The ninth paragraph summarizes the contributions of the paper: idea, relationship, performance, generalization.
·Methods/Experiments/Materials
Through this, we will know very clearly what the author is doing, and how is he going to do it? And what is the whole process of the specific experimental part. We will see how the author verifies his core ideas, and he further explains and refines his core ideas
·Results and discussion
·in conclusion
In this part, the author will review his core idea again, the methods and effects used, and possible improvement directions.
