# Labs

Seventeen Jupyter notebooks released across the semester in three chunks. Each lab is a self-contained exercise in a particular ML model type, loss type, or data type. The labs were originally distributed as Colab-runnable notebooks and were the substrate for three in-class lab days (L11, L12, L16).

The release schedule:

- **Chunk 1 — ML model types** (released **Sun Feb 22, 2026**, used for [Lab Day #1](../lectures/lecture-11.md) and [Lab Day #2](../lectures/lecture-12.md)).
- **Chunk 2 — ML loss types / objectives** (released **Mon Mar 2, 2026**, used in [Lab Day #3](../lectures/lecture-16.md)).
- **Chunk 3 — Data types / modalities** (released **Fri Mar 6, 2026**, used in [Lab Day #3](../lectures/lecture-16.md)).

Students worked the labs in order during the semester. **The labs in this book are the *solution* versions — every TODO is filled in and every code cell has run with its outputs visible.** Use them to study: read the prompts, then read the worked solutions, then run them in Colab to experiment.

Click any lab below to view it inline. To re-run it yourself, use the launch button at the top right of any notebook page (Open in Colab / Open in Binder).

> **Note.** For chunk 1 (labs 0-5) the solutions are a separate notebook — the version shown here. For chunks 2-3 (labs 6-16) the original released notebooks already include solution cells inline at the bottom of each section, marked "Solution: ...".

## What each lab covers

### Chunk 1 — ML model types

| Lab | Topic | Connects to |
|---|---|---|
| [Lab 0](lab0_preprocessing.ipynb) | Preprocessing, EDA, missing data, scaling, train/val/test splits | Foundation for every other lab |
| [Lab 1](lab1_logistic_regression.ipynb) | Logistic regression on the breast-cancer dataset | L9 (binary-classification training) |
| [Lab 2](lab2_decision_trees.ipynb) | Decision trees; interpretability vs. overfitting | L10 (capacity / regularization) |
| [Lab 3](lab3_knn.ipynb) | k-Nearest Neighbors; bias-variance via k | L9 (kNN consistency) |
| [Lab 4](lab4_ensemble_methods.ipynb) | Random forests and gradient boosting | L10 (variance reduction) |
| [Lab 5](lab5_neural_networks.ipynb) | First feedforward neural network | L13 (NN lecture) |

### Chunk 2 — ML loss types / objectives

| Lab | Topic | Connects to |
|---|---|---|
| [Lab 6](lab6_multiclass_classification.ipynb) | Going from binary to K-way; softmax + cross-entropy | L4 (information theory), L7 (NLL) |
| [Lab 7](lab7_regression.ipynb) | Squared error, regression metrics, prediction intervals | L7 (Gaussian NLL) |
| [Lab 8](lab8_survival_analysis.ipynb) | Cox proportional hazards as NLL with censoring | L23 (population health, survival) |
| [Lab 9](lab9_probabilistic_regression.ipynb) | Predicting distributions, calibration | L8 (calibration vs. discrimination) |
| [Lab 10](lab10_clustering.ipynb) | k-means, GMMs, the unsupervised side of L7 | L7 (mixture models) |
| [Lab 11](lab11_dimensionality_reduction.ipynb) | PCA, t-SNE, UMAP | L2 (eigenanalysis), L25 (population PCA) |

### Chunk 3 — Data types / modalities

| Lab | Topic | Connects to |
|---|---|---|
| [Lab 12](lab12_graph_data.ipynb) | Graph neural networks, message passing | L26-L27 (molecules, biological AI) |
| [Lab 13](lab13_ordinal_sequences.ipynb) | Sequence inputs; RNN/transformer architectures | L14 (LLMs) |
| [Lab 14](lab14_images.ipynb) | Convolutional networks on images | L21-L22 (medical imaging) |
| [Lab 15](lab15_timeseries.ipynb) | Irregular sampling, missingness, time-series losses | L17-L18 (EHR data) |
| [Lab 16](lab16_event_streams_MEDS.ipynb) | Health event streams in MEDS format | L17-L18 (EHR data) |

## How to run the labs

Each lab notebook has Colab-friendly boilerplate at the top. To run a lab:

1. Open the notebook page in this book (links above).
2. Click the launch icon at the top right of the notebook page; it opens the notebook in Colab with one click.
3. Run cells in order.

### Runtime expectations

Most labs run comfortably on a free Colab CPU runtime in under 10 minutes. The exceptions:

| Lab | Approx. runtime | Free CPU OK? | Free GPU helpful? | Notes |
|---|---|---|---|---|
| Lab 0-5 | 1-3 min each | yes | no | Pure scikit-learn / numpy / matplotlib. |
| Lab 5 (NN) | 3-5 min | yes | no | Small NN; no GPU needed. |
| Lab 6-9 | 1-3 min each | yes | no | Loss-type variants; classical methods. |
| Lab 10-11 | 2-5 min each | yes | no | Clustering / dim-reduction; UMAP install can take ~30s. |
| Lab 12 (graphs) | 5-15 min | yes | recommended | Installs `torch_geometric` and trains a small GCN; faster on GPU. |
| Lab 13 (sequences) | 10-30 min | acceptable | **recommended** | Fine-tunes a small transformer; CPU works but is slow. |
| Lab 14 (images) | 10-30 min | acceptable | **recommended** | CNN training on MedMNIST; downloads ~80 MB of data. |
| Lab 15 (time series) | 5-15 min | yes | no | `wfdb` data downloads can take a minute. |
| Lab 16 (MEDS) | 5-15 min | yes | no | Downloads small MIMIC-IV demo data in MEDS format. |

### Cross-lab dependencies

Labs 0-11 share a `processed_data.pkl` produced by Lab 0. **Run Lab 0 first** (or download its output) before opening labs 1-11. Labs 12-16 are independent of that pickle but each needs its own dataset download (graph data, sequence corpus, MedMNIST, ECG signals, MIMIC-IV demo respectively).

If you're reading this book locally, you can also open any lab directly in Jupyter / VS Code from `Labs/notebooks/`. The Python environment used to render this book in CI is documented in `pyproject.toml` (Jupyter Book itself); per-notebook scientific dependencies are documented in `scripts/notebook_extras.tsv` and are installed into ephemeral isolated venvs by `scripts/execute_notebook.sh` during book builds.