Teaching · Fall 2026–2027
AIE 683 Machine Learning
This is a graduate course on machine learning with an emphasis on doing it rigorously and running it in practice. It covers the foundations of statistical learning; the main model families (linear models, tree ensembles and gradient boosting, neural networks); the practice of evaluating, tuning, comparing and interpreting models; unsupervised learning; and time series forecasting. Machine learning is treated as a living process rather than a one-off notebook: experiments are tracked with MLflow throughout the term, and the final weeks cover the ML lifecycle of registering, serving, monitoring and retraining models. Students carry out an empirical ML study and write it up as a research paper.
- Lectures
- Thursdays 08:40–11:40
- Office hours
- Thursdays 12:00–13:00 and 16:00–17:00
- Programme
- MSc Artificial Intelligence, Graduate School of Informatics
- Language
- English
Learning outcomes
- Explain the core concepts of statistical learning: generalization, the bias–variance trade-off, overfitting, and regularization.
- Build end-to-end ML pipelines in Python, from preprocessing through model selection.
- Choose algorithms suited to a given problem and data type, and justify the choice.
- Evaluate models rigorously with appropriate metrics, resampling, and statistical comparison.
- Recognize and avoid common pitfalls of applied ML, such as data leakage and distribution shift.
- Track, version, deploy and monitor models across the ML lifecycle using MLflow.
- Design, run, and report an ML study as a research paper.
Weekly plan
| Week | Topic and readings | Slides | Due |
|---|---|---|---|
| 1 | Introduction: ML problem framing, the ML workflow; tooling (uv, Git, MLflow)IMLP Ch. 1 · AML: Introduction | slides · PDF | |
| 2 | Generalization & validation: nearest neighbors, bias–variance, cross-validation strategiesIMLP Ch. 2 (k-NN), Ch. 5.1 · AML: Supervised learning, Data splitting strategies | slides · PDF | |
| 3 | Preprocessing, feature engineering, pipelines, data leakageIMLP Ch. 3.3, Ch. 4, Ch. 6 · AML: Preprocessing, Pipelines · Kapoor & Narayanan (2023), Leakage and the reproducibility crisis in ML-based science, Patterns | slides · PDF | HW1 out |
| 4 | Linear models: regression, regularization, logistic regression, SVMsIMLP Ch. 2.3.3, 2.3.7 · AML: Linear regression, Linear classification | slides · PDF | |
| 5 | Tree ensembles: decision trees, random forests, gradient boostingIMLP Ch. 2.3.5–2.3.6 · AML: Trees, Random forests, Gradient boosting · Grinsztajn, Oyallon & Varoquaux (2022), Why do tree-based models still outperform deep learning on typical tabular data?, NeurIPS | slides · PDF | HW1 due (22 Oct) · HW2 out · Paper proposal (extended to 29 Oct) |
| 6 | Neural networks (MLP): backpropagation, regularization, MLPs on tabular dataIMLP Ch. 2.3.8 · AML: Neural networks | slides · PDF | |
| 7 | Evaluation: metrics, calibration, imbalanced data, decision thresholdsIMLP Ch. 5.3 · AML: Evaluation metrics, Calibration, Imbalanced data | slides · PDF | HW2 due (12 Nov) · HW3 out |
| 8 | Model selection & experiment tracking: tuning, nested CV, model comparison, MLflow TrackingIMLP Ch. 5.2 · AML: Parameter tuning & AutoML · Nadeau & Bengio (2003), Inference for the generalization error, Machine Learning · Demšar (2006), Statistical comparisons of classifiers over multiple data sets, JMLR · MLflow Tracking docs | slides · PDF | |
| 9 | Interpretation & feature selection: permutation importance, PDP, SHAPIMLP Ch. 4.5 · AML: Interpretation, Feature selection | slides · PDF | HW3 due (26 Nov) |
| 10 | Unsupervised learning: PCA, t-SNE/UMAP, clustering, mixture modelsIMLP Ch. 3.4–3.5 · AML: Dimensionality reduction, Clustering | slides · PDF | Progress report (3 Dec) |
| 11 | Time series forecasting: lag features, time-based CV, boosting vs. classical baselinesHyndman & Athanasopoulos, Forecasting: Principles and Practice, 3rd ed. (free online), Ch. 1, 5, 8 | slides · PDF | HW4 out |
| 12 | The ML lifecycle: MLflow Models & Registry, serving, monitoring, drift, anomaly detection, retrainingSculley et al. (2015), Hidden technical debt in machine learning systems, NeurIPS · AML: Outlier detection · MLflow Model Registry docs | slides · PDF | |
| 13 | Project presentations | HW4 due (24 Dec) · Presentation (slides 23 Dec) | |
| Finals | Final paper & code |
The plan may shift with the pace of the class; changes are announced in class and on GitHub.
Assessment
| Component | Weight | Due |
|---|---|---|
| Paper proposal | 10% | 29 Oct (extended from Week 5) |
| Homework assignments (4 × 5%) | 20% | Weeks 5, 7, 9, 13 (22 Oct, 12 Nov, 26 Nov, 24 Dec) |
| Project & research paper | 70% | Progress report Week 10 (3 Dec); presentation Week 13 (24 Dec; slides due 23 Dec); final paper in finals week (date TBA) |
All deadlines are 23:59 Turkey time on the date shown. Homework is applied and graded automatically; the project is an original ML/DL research study with an academic contribution, written up as a research paper. Details: homeworks · project and proposal rubric.