Thesis/paper structure
Always structure your thesis or research paper according to the general convention. This ensures your reader knows where to look for certain information.
Intended for: BSc, MSc, PhD
A thesis or research paper needs to contain the following sections, which we discuss in more detail below:
Abstract
Length: 10-15 sentences.
Content: Your abstract should be short (e.g. 10-15 sentences), and summarize the entire thesis/paper. Cover (1-3 sentences per item):
The topic/setting.
The (remaining) problem/challenge.
Your solution.
The obtained results.
What this adds / why this is relevant to the reader.
Introduction
Length:
Paper: ~1 page (3-5 paragraph)s. One paragraph per topic, you may merge items 1/2 and/or 3-4 into a single paragraph.
Thesis: 2-5 pages.
Content: The introduction should explain to the reader whether or not this is a paper/report/thesis to further read (it is essentially a long version of the abstract). Cover:
Topic/setting: what type of problem are you working on, and why is this a relevant problem?
Problem identification: what is the state-of-the-art approach for this problem (cite the most relevant related literature), and what is still lacking in the state of the art?
Method: what do you propose to solve this deficit?
Results: What results do you find? What does your paper add to the literature, or teach the reader?
Contribution: Try to explicitly list the novel contributions of your paper.
Structure of document (optional): describe the overall structure of your document
Related Work
Length:
Paper: ~0.5-1 page. Thesis: 1-2 pages.
Note: You may also position this section just before the Discussion (after Results).
Content: Related work is an important aspect of a scientific report. This page describes how to find related work.
Structure: Often, your research has different connections, e.g., there are different groups of approaches to your problem, or your research combines ideas from multiple research fields. For each group, write a separate paragraph in which you group related work based on commonalities.
Compare: After you describe what another paper has done, you should also compare it to your work. Example: "Compared to method X, we make procedure Y stochastic by injecting noise in operation Z."
Background/Preliminaries (optional)
Length:
~0.5 page
Content: This section is optional. It should establish common ground for the rest of your report to build upon, and thereby keep the report self-contained (notation wise). Consider:
The problem setting/definition. You may want introduce the Markov Decision Process formulation, or the principles of Bayesian learning, or deep learning, etc.
Common notation. Very importantly, along with the above you should also introduce your basic notations and symbols you will use throughout. Never use a symbol which you have not first defined!
Existing standard algorithms/methods which you will use/extend in this work.
Methods
Length: ~2-4 pages.
Content: Write a methodology section, in which you 1) describe the algorithm/method you studied, and 2) detail the experimental set-up.
Algorithms
Fully describe what you implemented, in such a way that a reader could reproduce your work from the text.
Use equations and algorithm boxes (pseudocode) to detail the computational procedure (you can’t fully explain this in words).
You may use a figure to visually illustrate what happens in your method (can be very helpful).
Experimental set-up
Describe the environments/tasks/datasets you will use to study your algorithm. (e.g., describe the dimensionality of state and action space, provide a figure of the task, etc.).
Mention the hyperparameter values you used.
Once you start listing them, you usually find out there were more than you realised.
Describe which hyperparameters you tuned/optimized, and how you did this (e.g., ‘for the learning rate we performed a grid search over the range [0.001, 0.01, 0.01]’)
Attach your source code to reproduce the results, or provide a link to a public repository.
Results
Length: ~2-4 pages
Content: When generating results, simply keep asking yourself: what do I want to learn next about this problem and the possible solutions. Afterwards, you summarize the outcomes of this process to the reader. Think about:
Which aspects of your algorithm are relevant to vary.
Which performance measures you want to use, and how to display/visualize these.
Which other visualizations may provide additional insight in what is happening in your method.
Afterwards, write a results section:
Group your results per topic/(sub) research question.
Use figures and tables to summarize your results.
Give every figure and table a caption. Keep the numbering ordered.
Captions should be self-contained (you should be able to understand the figure without looking at the main text). Describe the set-up of the figure, and what we learn from it.
In the main text, interpret your results (very important).
Indicate what you learn from a figure/table, come up with possible explanations, make a connection to a next figure/table where you further investigate your hypothesis.
Focus your results on the most important outcomes/insights.
Additional results can added in an Appendix, to which you briefly point in the main text.
Discussion
Length: ~1 page.
Content: In the Discussion you reflect on you work, primarily discussing:
Weaknesses: Reflect on you own work, discussing the weaknesses of your methods and design.
To what extend are these weaknesses limiting? Why did you make a certain design choice? Would you do anything different next time?
Future work:
Given your current conclusions, what would be an interesting direction to further extend your ideas.
Conclusion
Length: ~1 paragraph.
Content: In the Conclusion you summarize the overall story of your report. Describe in one paragraph:
What method you proposed/studied.
What results you found.
Why this is useful/relevant to the reader.
Note: You may also merge the Discussion and Conclusion together into a single section (which you either call “Discussion” or “Conclusion”).
References
Length: ~1-... pages.
Content: References connect your work to other literature, and assign credit. Visit this page for a detailed discussion of references. In short:
No bullet points: Do not make references a bullet point list. Instead, either sort them on last name, and cite with (author, year) in the text, or give them a number, and cite in the text with [4].
Use a reference manager (such as bibtex): This will automatically generate your references and reference style, provides a good lay-out, and save you an extreme amount of hassle.
Make references clickable: Use the hyperref package to make all your references clickable. Simply enter: \usepackage{hyperref} to your document preamble.
Note: The reference section is not numbered. Latex automatically handles this when you generate the appendix from a bib file with the \bibliography{...} command.
Appendix
Length: ~1-20 pages (technically as many as you need).
Content: You use the appendix for elements of your text that take much space, and would thereby intervene in the flow of the report, but are nevertheless interesting for readers that want to dig deeper. Consider:
Full experimental details: like a full table of all hyperparameters.
Additional results: for example extra visualizations of results. This quickly takes up much space.
Mathematical details: proofs or expansions that would take too much space in the main text.
Note: Appendix sections are usually numbered from A-Z. Enter the command \appendix in your Latex document after your references, and the next section will start at 'A'.