This repository contains guidance for students doing a master's thesis supervised by Alex Jung at Aalto University.
- Currently supervised theses: theses.md
- Available topics: Topics.md
- Official SCI thesis guide (templates, forms, evaluation rules): SCI Master's Thesis Guide on MyCourses — start with the "Get oriented" section
- What to Expect from Your Supervisor
- What Is Expected from You
- Getting Started
- Typical Timeline
- Recommended Development Environment
- Responsible Use of AI
- Practical Workflow
- Thesis Manuscript Preparation
- Typesetting Mathematical Texts
- Iterative Writing Process
- Self-Editing Pass (Prose Linter)
- Final Thesis Checklist
- Thesis Presentation and Self-Evaluation
- Thesis Evaluation, Decision, and Appeals
- References
- Feedback and Questions
As your supervisor, you can expect me to:
- Help you clearly define your ML problem.
- Advise on suitable ML methods, tools, and resources.
- Guide you through thesis writing and evaluation.
- Provide feedback on drafts and self-assessments.
- Offer the opportunity to discuss your self-assessment before submission.
As a thesis student, you are expected to:
- Take ownership of your research and drive progress independently.
- Come to meetings prepared with concrete questions or results.
- Communicate proactively — notify me early if you are stuck or falling behind schedule.
- Use high-quality scientific references (peer-reviewed journals, reputable conferences, established textbooks).
- Follow the writing and typesetting conventions described in this guide.
To start your thesis:
- Formulate your ML problem clearly by identifying data points, their features, and labels (watch this video).
- Choose suitable ML models that you are comfortable implementing (e.g., linear regression, random forests, or neural networks).
- Identify data sources and evaluation criteria (e.g., test accuracy, computational efficiency).
Detailed guidance is available in Chapter 2 of the textbook and in these lecture videos.
A master's thesis typically spans 6–12 months:
| Phase | Activities |
|---|---|
| Problem Definition | Identify research question, data sources, and evaluation criteria |
| Literature Review | Survey related work; identify gaps your thesis addresses |
| Data Collection & Preprocessing | Gather, clean, and explore your dataset |
| Modeling | Implement and train ML models; run baseline experiments |
| Evaluation & Diagnosis | Benchmarks, sensitivity analysis, error analysis |
| Writing | Draft chapters iteratively; incorporate feedback |
| Self-Assessment & Presentation | Complete evaluation form; prepare and deliver thesis presentation |
Visual Studio Code is a good choice for thesis work — it supports LaTeX (via the LaTeX Workshop extension), Python, and Jupyter notebooks in one place.
The Claude Code extension for VS Code integrates an AI assistant directly into your editor. You can use it to:
- Explain or refactor Python/LaTeX code in context
- Get feedback on a selected paragraph or equation
- Generate boilerplate (e.g., plotting code, pseudocode skeletons)
- Ask questions about your codebase without leaving the editor
Note: You are accountable for everything in your thesis; an LLM cannot be. Verify every AI-generated output before you rely on it.
Follow Aalto University's official policies:
You are accountable for the entire content of your thesis — every claim, result, and conclusion. An LLM cannot be: it cannot be examined, cannot answer for a mistake, and cannot be listed as an author. "The AI produced it" excuses nothing. Whatever tools you use, you must be able to defend every statement in your thesis as if you had written it unaided.
- Disclose when and how you used AI tools, in a dedicated statement — not in the Methods section, which is reserved for research methods.
- Record the tool, version, and settings; online services change frequently, so exact reproduction is rarely possible.
- Do not upload personal, confidential, or unpublished material to public AI services — this may breach GDPR. Use local or GDPR-compliant tools for sensitive data.
- AI output may reproduce others' work without attribution. Verify every citation independently.
AI tools can support your work when used for:
- Proofreading language and grammar
- Brainstorming research directions or experiment designs
- Explaining concepts or summarising background (always verify)
- Generating boilerplate or plotting templates (always review and test)
- Surfacing counterarguments to test your reasoning
A master's thesis in machine learning typically involves:
- Data Collection and Preprocessing using, e.g., pandas.
- Model Training and Validation using, e.g., scikit-learn.
- Model Diagnosis using numerical experiments (benchmarks, sensitivity analysis) and, when appropriate, mathematical analysis (generalisation bounds, error analysis, comparison to Bayes' risk).
For academic sources, use:
- Aalto University Library: https://primo.aalto.fi/discovery/search?vid=358AALTO_INST:VU1&lang=en
- IEEE Xplore: https://ieeexplore.ieee.org
- ACM Digital Library: https://dl.acm.org
- Scopus: https://www.scopus.com
- Web of Science: https://www.webofscience.com
To assess the quality of journals and conferences, consult the JUFO ranking system: https://jfp.csc.fi/jufoportal
If you are uncertain about a reference's quality, ask me.
When preparing your thesis, ensure:
- Terminology: Use terms defined in the Aalto Dictionary of ML. You are encouraged to reuse its LaTeX source (e.g., TikZ figures).
- Problem formulation: State clearly what the data points are and how their features and labels are defined.
- Loss functions: Explicitly state the loss function used for training and, separately, for validation or testing.
- Numerical results: Present and discuss results thoroughly to answer your research questions.
- Baselines: Use appropriate baselines or benchmarks (e.g., Kaggle competitions).
- Structure: Begin each chapter and section with an introductory paragraph explaining its content and its connection to the rest of the thesis.
- Equations: Reference all numbered equations using
\eqref{}. Only number equations that are referenced in the text; leave unreferenced equations unnumbered. - Algorithms: Present new methods as pseudocode (see examples).
- Figures: Ensure all figures are clear, labelled, and have informative captions (caption guidelines).
- References: Format according to IEEE guidelines.
For creating effective figures, see Edward Tufte's The Visual Display of Quantitative Information.
- Use inline math (
$...$) for short expressions within a sentence:The loss is defined as $L(\theta)$. - Use display math (
\[ ... \]or theequationenvironment) for standalone equations that are central or referenced.
Punctuate displayed math as part of the surrounding sentence:
The empirical risk is defined as
\[
L(\theta) = \frac{1}{n} \sum_{i=1}^n \ell(f(x_i;\theta), y_i).
\]- While experimenting, keep a separate working notes file with the methods you try, the hyperparameters used, and the results — without worrying about prose, formatting, or citations. Treat it as a lab notebook.
- Start serious writing with the literature review or methodology chapters. These stabilise earliest and rarely need to be rewritten when results shift.
- Defer the results and discussion chapters until your experiments have converged — they often change substantially, and writing them up too early wastes effort.
- Write the abstract last. It is the shortest section but depends on everything else being settled.
- Incorporate feedback regularly from peers, group meetings, or, where appropriate, LLM-based tools.
- Expect and budget for multiple revision rounds before submission.
Once a draft is near-final, read it through once looking only for the recurring prose defects below. Each is cheap to fix yourself and distracting for a reader — or your supervisor — to flag repeatedly. Fixing them before review means feedback can focus on the substance of your work rather than on mechanics.
- Excessive forward referencing — "as we will see in Chapter 5" scattered throughout. Present material where the reader needs it; a few pointers are fine, but a habit of them signals disorganised structure. A small helper script,
forward_reference_linter.py, flags forward-looking cue phrases and any\ref/\eqrefwhose\labelis defined later in the document — runpython3 forward_reference_linter.py chapter1.tex chapter2.tex ...(files in reading order) as a first pass. - Undefined or re-defined acronyms — expand each acronym at first use, then use it consistently; do not re-expand it later.
- Inconsistent terminology — pick one term per concept (following the Aalto Dictionary of ML) and do not silently switch between synonyms (e.g., "data point" vs "sample" vs "instance").
- Vague quantifiers — "significantly", "very", "a lot" without a number. Give the figure or drop the word.
- Uncited claims — every empirical or historical assertion that is not your own result needs a citation.
- Dangling references — "this shows", "it follows" where "this" or "it" has no clear antecedent.
- Unmotivated sections — a chapter or section that opens without a sentence stating what it covers and why it belongs here.
- Tense drift — keep the tense consistent within a passage when describing methods and results.
Reading aloud, or reading one defect type at a time from start to finish, catches far more than a single general pass.
Before submitting, verify each item below. Links point to the relevant section of this guide.
- ML problem is precisely formulated: data points, features, and labels are clearly defined (see Getting Started)
- Loss functions for training and evaluation are explicitly stated (see Manuscript Preparation)
- Methods are clearly described, including pseudocode for new algorithms
- Baselines or benchmarks are included and discussed
- All figures have labelled axes and informative captions
- All numbered equations, tables, and figures are referenced in the text
- Citations are formatted according to IEEE guidelines
- Acronyms are expanded at first use and terminology is consistent (see Self-Editing Pass)
- A prose self-editing pass has been done for forward referencing, vague quantifiers, and dangling references (see Self-Editing Pass)
- Self-assessment form is completed (form here)
After completing the thesis manuscript:
- Complete the detailed self-assessment (evaluation form), with explicit references to sections of your thesis.
- Review the grade characterisation PDF to understand what constitutes a high-quality thesis.
- Optionally, request a meeting to discuss your self-assessment before submission.
- Prepare your thesis presentation — either live during a group meeting or as a recorded video (see examples).
- Your thesis is evaluated against the official programme and school-level criteria.
- I prepare a written evaluation and a grade proposal based on these criteria.
- The final grade is confirmed by the programme or school following Aalto University's formal procedures.
- Study the grade characterisation document carefully before submission.
- You have the right to see the evaluation criteria applied to your thesis.
- You may request clarification on how your thesis was assessed and how the grade was formed.
- The self-assessment form is an important part of this process.
- If you believe an error occurred, you have the right to request rectification of the grade.
- Appeals must be submitted within 14 days of being notified of the grade or of being given the opportunity to review the evaluation.
- See Academic Appeals at Aalto University for the formal procedure.
Practical advice: If unsure whether an appeal is appropriate, discuss the evaluation with your supervisor first. Many issues can be resolved without a formal appeal.
- A. Jung et al., "The Aalto Dictionary of Machine Learning," Aalto University. [GitHub]
- S. Shalev-Shwartz and S. Ben-David, Understanding Machine Learning: From Theory to Algorithms. Cambridge University Press, 2014.
- Peer-Review Form — Machine Learning Course Project
- Peer-Review Form — Federated Learning Course Project
- D. Bertsekas, Ten Simple Rules for Mathematical Writing
- D. E. Knuth, T. Larrabee, and P. M. Roberts, Mathematical Writing
- E. Tufte, The Visual Display of Quantitative Information, 2nd ed., 2001.
- AMS Style Guide for Journals
- IEEE Editing Mathematics Style Guide
- IEEE Editorial Style Manual for Authors, 2024
Reach out via:
- Email: alex.jung@aalto.fi
- LinkedIn: linkedin.com/in/aljung/
- YouTube: @alexjung111
- GitHub: Use issues or pull requests on this repository.