Tampere University Signal Processing and Machine Learning Master's Thesis Guide: 30 ECTS, COMP.SGN-S02 and Trepo

Quick answer: what governs the Signal Processing and Machine Learning thesis?

Tampere University’s Signal Processing and Machine Learning option is a 120 ECTS, two-year Master of Science (Technology) specialisation in Computing Sciences and Electrical Engineering. The current programme wrapper is TTEM-SPMLY, while the thesis-bearing engineering major is COMP.SGN-S02 Advanced Studies in Signal Processing and Machine Learning, worth at least 80 ECTS. Its public 2026-2027 structure is unusually explicit: six mandatory courses total 30 ECTS, free-choice studies contribute at least 20 ECTS, and the master’s thesis is 30 ECTS.

Because this is an MSc (Technology) route, the thesis follows Tampere’s Technology thesis process and is graded on the 0-5 scale. A parallel current module, COMP.SGN-S01, is at least 60 ECTS but does not embed the thesis in its public module description. Do not silently swap those structures.

1. The programme is broader than “do deep learning”

The applicant page names signal-processing theory and algorithms, deep learning and neural networks, audio and speech processing, imaging and computer vision, and statistical modelling/data analysis. COMP.SGN-S02 also points to imaging, machine hearing and vision, and signal processing/data analysis as recommended thematic areas.

That breadth matters. A thesis can centre on classical estimation, adaptive filtering, spectral methods, image reconstruction, audio processing, statistical modelling, computer vision, neural networks or a hybrid. Tampere’s current Advanced Signal Processing course still covers Wiener and least-squares methods, adaptive filters and spectrum estimation. A strong thesis chooses the method that fits the question rather than treating the newest model family as mandatory.

2. Read COMP.SGN-S02 as the thesis-bearing major

COMP.SGN-S02 is the clearest current major object for this guide because it explicitly includes the 30 ECTS thesis inside its at-least-80-credit structure. It is current through 2026-2027, taught in English and designed around adopting state-of-the-art methods for practical problems.

The parallel COMP.SGN-S01 is also current and has related learning outcomes, but its public structure is at least 60 ECTS and does not embed the thesis in the same way. Treat S01 as a curriculum control, not a reason to rewrite the MSc Technology major. Your own Sisu plan remains the operational authority for the exact studies attached to your study right.

3. A 30 ECTS Technology thesis needs an academic argument

A working model, successful reconstruction or high benchmark score is not yet a thesis argument. The thesis should define a research problem, explain why the chosen method is suitable, establish a comparison or evaluation protocol, analyse the evidence and discuss limitations.

Implementation can be central. You may build a training pipeline, signal-processing algorithm, imaging method or embedded inference system. But the academic contribution must be visible around the artefact. Readers should understand what claim is being tested, what evidence would support or weaken it, and where the result is expected to generalise.

4. Turn a broad AI idea into a bounded question

“Use deep learning for images” or “apply machine learning to audio” is a field, not a research question. Specify the input, output, population or signal conditions, comparator and target outcome. For classification, define classes and operating conditions. For enhancement or reconstruction, define degradation and quality criteria. For forecasting, define horizon and information available at prediction time.

A narrow question improves both engineering and writing. It determines what data are needed, which baseline is credible and which metric actually matters. It also prevents scope creep into data collection, architecture design, deployment and productisation all at once.

5. Build the data split before building the model

Machine-learning theses are especially vulnerable to data leakage. Decide how training, model selection and final testing are separated before repeated experimentation begins. The correct split unit may be a person, recording session, patient, speaker, device, site, time interval or physical specimen rather than an individual sample.

If multiple windows come from the same recording or subject, a random sample split can make test data statistically dependent on training data. If the research claim concerns unseen subjects or devices, split at that higher level. Preserve the exact split logic so the experiment can be reproduced.

6. Prevent preprocessing leakage

Leakage can occur before a model sees the data. Normalisation statistics, feature selection, dimensionality reduction, denoising parameters, augmentation policies or missing-value handling can accidentally use information from the final test set.

Fit data-dependent preprocessing on the training partition and apply the resulting transformation to validation and test data unless the research design explicitly justifies another approach. If preprocessing is part of the proposed method, evaluate it inside the same controlled protocol as the model.

For signal data, segmentation itself can leak information. Overlapping windows from one source should not be scattered across train and test if that violates the intended generalisation claim.

7. Keep the final test set genuinely final

Repeatedly checking the test set while tuning architecture, thresholds or features gradually turns it into validation data. Use validation or nested model-selection procedures for iterative decisions and reserve the final test set for the last evaluation.

If the dataset is small, cross-validation may be more appropriate, but the grouping rule still matters. Subject-wise, session-wise or site-wise folds can be essential. Report how folds were constructed, how hyperparameters were selected and whether the reported result aggregates across folds or comes from a final refit.

8. Choose baselines that make the claim meaningful

A new method needs comparison against something credible. Depending on the question, that might be a simple statistical baseline, classical filter, linear model, established neural architecture, previous published method or current production approach.

Use the same data split and evaluation protocol for all methods. If one baseline receives less tuning or weaker preprocessing, the comparison is difficult to interpret. Explain why each baseline is relevant: does it test whether complexity is necessary, whether signal-specific knowledge helps, or whether the proposed contribution improves the current state of practice?

9. Metrics should correspond to the actual task

Accuracy can hide failure on rare classes. Precision, recall, F1, ROC/PR measures, calibration or class-specific error may be more informative depending on the application. For regression, MAE, RMSE or domain-specific tolerances answer different questions. For reconstruction, image or audio quality metrics can disagree with perceptual usefulness.

Choose a primary metric before final evaluation when possible. Report supporting metrics when they reveal trade-offs rather than simply listing everything a library returns. If deployment has constraints, latency, memory, throughput or energy can be legitimate secondary outcomes.

10. Report uncertainty, not only the best run

Neural-network training, random data splits and stochastic optimisation can produce different results across runs. One favourable seed is weak evidence. Use repeated runs or appropriate resampling when variability matters and report mean/median plus spread or confidence intervals where suitable.

For deterministic signal-processing pipelines, uncertainty can still come from measurement noise, finite samples or estimated parameters. Explain what variation exists and how it affects the conclusion. A small numerical improvement may not be meaningful if it is smaller than normal run-to-run variation.

11. Signal-processing theses need explicit assumptions

Classical signal processing is not a fallback. Current Tampere content includes estimation, Wiener/least-squares filtering, LMS/NLMS/RLS adaptive methods and spectrum estimation. These methods rely on assumptions about signals, noise, stationarity, model order or convergence.

State those assumptions. If the signal is non-stationary, explain how window length or adaptation handles it. For spectral work, document sampling frequency, window, overlap, transform length and frequency resolution. For filtering, report design criteria, boundary treatment and stability or convergence considerations when relevant.

12. Audio and speech experiments

For audio or speech work, document sampling rate, channels, segmentation, loudness or amplitude normalisation, augmentation and acoustic conditions. Speaker leakage is a frequent validity problem: if the claim concerns unseen speakers, speaker identities should not cross the train/test boundary.

Noise conditions also define the claim. A method tested only on synthetically added stationary noise should not be described as robust to real-world acoustic variability without evidence. If using public corpora, record dataset version and licence conditions and describe exclusions or preprocessing clearly.

13. Imaging and computer vision experiments

Imaging work can be sensitive to acquisition device, site, resolution, compression, lighting, optics and preprocessing. If several images come from the same patient, subject, scene or video, split at the level that matches the generalisation claim.

Augmentation should be plausible for the data-generating process. Horizontal flipping may be harmless in one task and scientifically invalid in another. Document resizing, cropping, colour transforms and normalisation. If a model appears to exploit shortcuts or acquisition artefacts, investigate that possibility rather than assuming the network learned the intended phenomenon.

14. Time-series and forecasting work

Temporal prediction requires strict causality. Features, scaling statistics or labels from the future must not influence earlier predictions. Random shuffling can make a forecasting problem unrealistically easy when the intended deployment predicts future periods.

Use chronological or rolling evaluation when appropriate. Describe prediction horizon, look-back window and retraining schedule. If concept drift is possible, distinguish performance under the historical test period from claims about long-term future behaviour.

15. Hyperparameter search is part of the method

Architecture depth, learning rate, regularisation, augmentation, filter order and many other choices can materially change results. Record the search space and selection criterion. A thesis that reports only the winning configuration hides part of the research process.

Avoid giving the proposed method far more optimisation effort than baselines without acknowledging it. If compute limits constrain the search, state that. For automated optimisation, preserve random seeds and software versions where feasible.

16. Reproducibility: freeze the computational environment

Record dataset version, split files, preprocessing code, random seeds, model configuration, optimiser settings, training epochs, early-stopping rule and important library/framework versions. For Matlab or Python pipelines, version the scripts and environment definition.

If GPU nondeterminism or distributed training affects exact replication, say so. Reproducibility does not always mean bit-identical outputs; it means another researcher can reconstruct the method and understand expected variation. Keep experiment identifiers so plots and tables can be traced back to exact runs.

17. Robustness and domain shift

Strong benchmark performance does not automatically survive new microphones, cameras, scanners, speakers, patients, robots, noise conditions or geographic sites. Decide whether the thesis studies in-distribution accuracy or robustness across a meaningful shift.

If robustness is part of the claim, define the shift and test it deliberately. If it is not, state the boundary. Avoid broad wording such as “works in real-world conditions” when the evidence covers only one curated dataset.

18. Statistical significance and practical significance

With many models and metrics, some differences occur by chance. Use statistical tests only when their assumptions fit the design, and avoid treating a p-value as the size or importance of an effect.

Also consider practical significance. A 0.2-point metric gain may be irrelevant if it doubles latency or requires much more labelled data. Conversely, a modest average improvement may matter greatly on a safety-critical rare class. Connect numerical results to the application context.

19. Human or sensitive data

Speech, images, biomedical signals and behavioural data can involve personal data even when names are absent. Discuss controller roles, lawful handling, minimisation, access control, storage, retention and publication with the supervisor before processing sensitive material.

Dataset licences and consent conditions matter too. A dataset being downloadable does not automatically permit every reuse or redistribution. If participants are recruited, ethics review or informed-consent requirements may depend on the study design. Do not assume an engineering degree removes those obligations.

20. Company and research-group theses

Tampere encourages research-group integration, and industrial collaboration is common. A company dataset or model can be valuable, but academic evaluation still needs transparent reasoning and sufficient evidence.

Define what can be published before the project becomes dependent on confidential material. The thesis itself is a public document. Proprietary code or data can remain outside the manuscript, but the public thesis must still explain the method, evaluation and limitations well enough for examination.

Ablation studies should answer a specific mechanism question

When a proposed pipeline contains several new elements, an ablation can help identify which component actually drives the result. Remove or replace one element at a time where that comparison is scientifically meaningful. Do not build a large ablation table merely because machine-learning papers often contain one. Each comparison should correspond to a claim about architecture, preprocessing, loss, feature representation or signal-processing stage.

Keep the evaluation protocol fixed across ablations. If a smaller model is retrained with different data, augmentation or optimisation effort, the observed difference cannot be attributed cleanly to the removed component. When components interact strongly, acknowledge that one-at-a-time ablation may not isolate the full mechanism.

Compare computational cost fairly

If the thesis claims efficiency, report cost under comparable hardware and software conditions. Parameter count alone does not establish inference speed; FLOPs alone do not establish energy use. Depending on the claim, measure latency, throughput, peak memory, model size, training time or energy using a documented environment. Include batch size, precision, accelerator, framework and warm-up strategy where they materially affect the number.

For embedded or edge-oriented work, desktop-GPU results may not represent the target platform. If deployment evaluation is outside scope, say so rather than translating server benchmarks into unsupported edge-device claims.

Calibration and error analysis can matter more than another model

A classifier can have good average accuracy while producing poorly calibrated probabilities or systematic errors on important subgroups. If decisions depend on confidence, consider calibration or threshold behaviour. Examine confusion patterns, difficult examples and subgroup performance when they are relevant to the research question and ethically appropriate.

Error analysis should remain evidence-driven. Avoid selecting a handful of visually interesting failures and treating them as representative. Define categories or sampling rules, then connect the observed errors to data quality, model assumptions or domain conditions.

Maintain a claim-to-evidence table

Before final writing, list each major thesis claim beside the experiment, table, figure or analysis that supports it. Mark claims that are only hypotheses or engineering observations. This exposes unsupported conclusions early and prevents the discussion from becoming broader than the evidence. It also makes the final manuscript easier to structure: methods generate evidence, results present it, and discussion interprets exactly what that evidence can and cannot establish.

21. ITC.CEE.800 and the seminar boundary

Current ITC.CEE.800 Tools for theses is 5 ECTS and pass/fail. Its current four components are research methods, good scientific practice/research ethics, information searching and career planning. Tampere recommends completing the course before or while planning the thesis.

Older ITC.CEE.800 implementations included a one-credit thesis-presentation/seminar component and even named an SPML seminar. That is historical evidence, not a current 2026-2027 requirement. The current university Technology-thesis process still includes seminar, presentation and opposition activity, but use current Sisu/Moodle instructions for the operational implementation.

22. Supervision plan and thesis schedule

Technology students use a Thesis Supervision Plan. Agree on the research question, supervisor roles, meeting cadence, expected deliverables, data access, compute resources and target completion period early.

For ML work, reserve time for data cleaning and failed experiments. For signal-processing or imaging work, reserve time for acquisition, calibration or algorithm validation. Build milestones around evidence generation rather than only implementation: frozen dataset/split, baseline complete, primary experiment complete, analysis complete, manuscript draft and final checks.

23. AI use in an AI-heavy programme

Studying machine learning does not exempt the thesis from Tampere’s AI rules. AI tools can assist coding, debugging, idea generation or language improvement within current guidance, but you remain responsible for correctness and acknowledgement.

Do not upload confidential company data, unpublished datasets, personal data or protected code to external AI tools without permission. Verify generated citations, mathematical claims and code. AI output is not a substitute for primary literature or your own data analysis.

24. Maturity test, Turnitin and Trepo

The master’s degree includes a maturity test. Under Tampere’s current Technology guidance, the thesis abstract normally serves as the maturity test for international master’s students, while language-history circumstances can change the exact route.

After supervisor permission, the final manuscript goes through Turnitin originality checking. Similarity percentage is not an automatic plagiarism verdict; the report requires academic interpretation. The final thesis is submitted through Tampere’s Trepo process and must meet the required file and metadata conditions.

25. Publicity, PDF/A and confidential material

Approved master’s theses are public documents and are electronically archived. Prepare a valid PDF/A file and test conversion before the deadline, especially if the thesis contains many equations, figures or embedded fonts.

Keep confidential code, business data, credentials and restricted datasets outside the public manuscript. Open-web access and legal public status are not identical, but confidentiality cannot be achieved by placing secret material into the assessed thesis and asking the university to hide it afterward.

26. Assessment, finality and graduation

The MSc (Technology) thesis uses the 0-5 grading scale. This 0-5 thesis scale is separate from the pass/fail grading of ITC.CEE.800. ITC.CEE.800, if present in your plan, is a separate pass/fail course and does not determine the thesis grade. Tampere appoints examiners and uses its formal assessment criteria; the ordinary examiner window is generally 21 days after the relevant final submission stage, with 28 days where the separate maturity-test route requires it.

After approval the thesis is final rather than freely resubmittable. Leave administrative time for Trepo processing, maturity completion, examiner assessment and degree application. A technically finished model is not the same as a completed degree process.