8DM60 · Exam guide#
One short page per topic. Every idea is explained in plain words and ends with a question — answer it before you open the answer.
If you only have two hours#
| Time | Do this |
|---|---|
| 0–20 min | Cram sheet — the whole course on one page. |
| 20–40 min | Multiple choice · practice — 20% of the grade and the traps are predictable. |
| 40–90 min | Past exercises — cover the answer, say yours out loud, then check. |
| 90–110 min | The six topic pages below, skimming the answers. |
| 110–120 min | Clinical workflow cases — the numbers. |
Pass conditions#
| Component | Weight |
|---|---|
| Multiple-choice questions | 20% |
| Group-specific project questions | 15% |
| General open/problem-based questions | 50% |
| Project poster + code | 15% |
Written exam ≥ 5.5 · final grade ≥ 6. Lecture 7 is not exam material.
The six topics#
Each page: explanation → one question → hidden answer.
- Language models — tokens, attention, training, prompting, RAG, LoRA.
- Vision models — DINO, CLIP, VLMs, validation.
- Omics models — scVI, scGPT, CPA, GEARS, perturbation metrics.
- Models for hospital data — representations, bias, leakage, calibration.
- Clinical workflow integration — deployment, thresholds, governance, money.
- Agentic AI — tools, ReAct, workflows versus agents.
Extras#
- Key figures — the ten diagrams and the one line to say about each.
- Your project · 15% — how to answer the poster questions.
The one idea behind the whole course#
A model produces a score. A clinical decision needs an operating point, calibration, validation on local data, and a named owner for what goes wrong. Most exam questions are asking which of those is missing.