本文内容来自于周志华老师在2007年的报告《做研究与写论文》
首先讲到研究与研发的区别,其主要区别在于“新”:
- 研究:发现新知识、发明新技术
- 研发:根据已有知识和技术进行研制、开发
科学研究可以扩展人类的知识,没有科学研究就没有技术进步
如何做研究?
研究的基本过成可以分为TPIC,即Topic -> Problem -> Idea -> Concrete work (theoretical analysis, experiments).
1、Topic
计算机科学发展到今天,已经是一个非常广袤的学,先要进入一个具体的分支学科和领域,并获 得必要的了解(如NLP,CV等)。即使在一个分支学科和领域中,也有太多的话题(如NLP中的文本分类,文本摘要,预训练模型等)。
热点分布与可能对应的研究现状如下图所示:
- 这只是目前CS中的一般情况,国内往往滞后一段时间
- 科学发展往往是螺旋式上升的:“冷”了很久的一个topic,可能由于新的经典工作而重新“热”起来
2、Problem
“问题”是科学研究的心脏 !
- 任何有价值的研究,都是为了解决某个问题
- 提出一个好的问题,已经成功了一半
“问题”其实才是研究的真正开始
- 这可能是CS研究中最困难的部分
- 会找问题,是具有独立研究能力的标志
3、Idea
你也许会有很多idea,导师可以帮助你判断、改善 有了好的idea,问题就解决了一大半;
如果一下想不到好的idea,不要着急;
博士期间能做出一项重要的工作就已经很好了。
4、Concrete work
Idea需要得到支持 CS里面通常是理论分析和实验验证。
理论分析往往需要比较好的数学功底,实验验证通常需要较好的实验设计能力。
没有相应的能力怎么办?
两者都不容易,到博士念完,至少在某方面应该是得心应手了。
理论分析
实验验证
- 实验方案周全仔细
- 基准测试
- 其他学者也能使用的数据
- 不可缺少的比较
- 实验是可重复的
必要的分析和解释
- Strength/weakness
- How/When strong?
- Why strong?
- How/When weak?
- Why weak?
小结一下
Topic -> Problem -> Idea -> Concrete work
对一项具体的CS研究工作来说:
- 找到好的问题是关键
- 有了好的idea,问题就解决了一大半
- 具体、坚实的工作是必经的过程
如何写论文?
你需要说的其实就是这些 :
- Problem X is important
- Previous works A, B, and C have been done
- A, B, and C have their weakness
- Your work D
- Theoretic analysis
- Experimental comparison against A, B, and C
- Why D is better
- Strength and weakness of D
- Future work on D
Outline
科学论文有比较清晰的结构,一般包括 :
- Title
- Abstract
- Introduction
- Previous work
- Your contribution
- Support (theoretical or experimental) – Discussion
- Conclusion
- Reference
Title
清楚地表达出主要工作
An example:
- “甲地区乙时期丙昆虫交配过程的跟踪研究及其结果”
- “昆虫交配后吃掉配偶的原因探讨”
- “昆虫求偶过程中的献身行为”
- “Eat me!”
- The title of an <Science> paper
Abstract
Purpose: Summarize your contributions
Style:
- What is the problem
- What is your work
- Features of your work
- Advantages of your work
- Results
Introduction
Purpose: Background and organization of the paper
Style:
- Problem X is important
- A, B, and C have been done
- A, B, and C have their weakness
- Our work D
- Features and advantages of D
- Results
- Organization of the paper
Previous work
Purpose: Why your work, the differences
Style:
- Categorization of previous works
- One or two sentences for a work
Don’t over-criticize previous works
Your contribution
Purpose: Introduce your work
Style:
- Motivation
- Definition, notation
- Algorithm
- Pseudo-code
- Diagram
- Explanations
If you were the reader, what questions will you ask?
Theoretical analysis
Purpose: Theoretical support to your work
Style:
- Definition, notation (can be omitted if exist)
- Lemma
- Theorem
- Proof
Put tedious details in Appendix
Experiments
Purpose: Experimental support to your work
Style:
Put tedious details in Appendix
Discussion
Purpose: The relationship between your work and some very related works
Style:
- Work A
- Why it is very related
- Difference to your work
- Work B
- Why it is very related
- Difference to your work
Conclusion
Purpose: Summary and future work
Style:
Acknowledgement Reference Appendix
Misconceptions
Misconceptions 1:
The more, the better
Don’t provide too many “new” ideas in a paper. Try your best to focus on your core work
Misconceptions 2:
The more complex, the better
Try to make your paper easy to be understood.
Remember: You are showing your thoughts instead of confusing the readers
Mathematical language is to help you clearly express your thoughts, not to show your depth
Misconceptions 3:
The more selling, the better
Don’t exaggerate too much on your work. The reviewers/readers will judge it.
Misconceptions 4:
The more authoritative, the better
Don’t refer too many own work while ignoring others
Don’t behave as a bigshot if you were not
Tricks
-
Read more, write more
-
Use good papers as samples to simulate but not to copy
-
Use simple sentences
-
Not to show your elegance unless you are very famous
-
Try to avoid syntax errors
Good work won’t be rejected for syntax, but at least the paper should be readable
-
Make your abbreviations easy to be pronounced
- e.g. STARE is better than SBRE “STAtistics-based Rule Extraction”
-
Typeset according to the requirements of the target
-
Revise, revise, and revise
- each time lay your paper aside for some time ask other persons to read your paper
-
Improve from comments
总结
论文是为研究工作服务的
- 为了论文而论文,若非生计所迫, 就是浪费生命
- 在“高档”期刊会议上发表10篇被引用0次的 论文,远不如在“低档”期刊会议发表1篇被 引用10次的论文
古今做学问者必经的三个境界:
- 昨夜西风凋碧树,独上高楼, 望尽天涯路
- 衣带渐宽终不悔,为伊消得人憔悴
- 众里寻他千百度,蓦然回首,那人却在灯火阑珊处
学问苦乐,日久自悟