Transverse project (experience-based learning)
Transverse learning · Year 2 · 8 ECTS · mandatory
Contents
- Stakes (Extrapolating)
- Place in the 2-year path
- Industry and research links
- How to approach it
- Calendar, deliverables and rubric
- Trust work package
| Module | Transverse project (experience-based learning) |
| Period | M2 · transverse learning |
| ECTS | 8 |
| Status | Mandatory · project |
Stakes (Extrapolating)
Extrapolated notes. The transverse project (8 ECTS, experience-based learning) is where several master’s building blocks meet a real or realistic problem: messy data, user constraints, a shared deliverable. Empirically, it fights silos: not only “pass course X”, but make a system / study hold end-to-end.
Place in the 2-year path
In M2 beside the ethics seminar, humanities/languages/sports, and before or alongside the final internship. It capitalizes on M1 P1–P2 + research internship + M2 courses. A good dress rehearsal before the 24 ECTS internship.
The end-of-year projects lead into this module from both years. Each Year 1 project, done at the end of Period 2, has a Year 2 sequel (continuation table), so a team can carry its repository, evaluation kit and open questions forward. The five Year 2 projects are written to become this module’s work: take one as it stands, or keep its calendar, trust work package and rubric for a subject of your own.
Industry and research links
Industry: project mode (backlog, milestones, review), integrating ML/LLM into a product. Research: pilot study, reproducible prototype, early sketch of a paper or open-source repo.
How to approach it
- Define “done” in usage terms, not only model metrics.
- Document decisions and technical debt — reusable for the internship report.
- If the topic is Graph × LLM, anchor it in a Year 2 project brief and in the chapters of the source map below. The roles such a topic leads to are listed under After the master on the home page.
Subject bank
Five briefs fit this module. They are proposals of these notes, given in full on the projects page with their questions, milestones, rubrics and solo or group variants. Every Year 2 course outside the core is slot TBC, so each brief names fallbacks for the ones it depends on most.
Y2.1 Biomedical QA with calibrated abstention. Graph × LLM route, healthcare. Courses: LLM ch. 7 and ch. 9, Geometric Data ch. 2–3, Advanced GNN 8.3, Privacy & UQ ch. 9. Data: STaRK-Prime over the PrimeKG knowledge graph. Baseline: hybrid BM25 + dense retrieval (LLM 7.2). Success: Hit@5 significantly above the baseline on relational queries, and an error rate on answered questions at or below a target (for example 10 %), at a reported coverage. Fallbacks: without Advanced GNN, the Graph ML ch. 5 encoders; without Privacy & UQ, its ch. 9 as self-study.
Y2.2 Text-to-3D-molecule copilot. Graph × LLM route, molecules and pharma. Courses: Geometric Data ch. 7–8, LLM ch. 4 and 8.1, Graph Generative Models ch. 5–7, Control ch. 3, Privacy & UQ ch. 9. Data: QM9, with ChEBI-20 for text conditioning. Baseline: unconditional samples of a released equivariant diffusion checkpoint (EDM). Success: at least 95 % of 100 requests parsed into a valid specification, and property-interval coverage within 5 points of nominal on generated molecules. Fallbacks: for Graph Generative Models, Graph ML ch. 9 and Multimodal GenAI ch. 4; Control ch. 3 and Privacy & UQ ch. 9 as self-study.
Y2.3 Speak, see, shop. AI-systems route, recommendation and e-commerce. Courses: Speech Technology ch. 4–5 and ch. 10, Geometric Data ch. 6, LLM ch. 6–8, MLOps ch. 3. Data: Amazon Berkeley Objects (ABO) and its 3D models, Common Voice, MUSAN noise. Baseline: typed queries with BM25 + dense retrieval over titles and attributes. Success: nDCG@10 of spoken queries within an agreed margin of typed ones, and a P95 time-to-first-audio of at most 1.5 s on one GPU. Fallback: without Speech Technology, use Whisper and VAD as tools and budget latency with LLM 6.5 and MLOps ch. 3.
Y2.4 The natural-language dispatcher. AI-systems route with a Graph × LLM option, last-mile logistics; it continues Y1.2. Courses: Geometric Data ch. 7, LLM 8.1 and ch. 6, Control ch. 3, Advanced GNN 8.2, MLOps ch. 3 and ch. 8. Data: the Amazon Last Mile Routing Research Challenge routes, CVRPLIB, NL4Opt. Baseline: the PyVRP solver. Success: on 100 dispatcher requests, an agreed formalisation accuracy with zero silent errors, and route cost within an agreed percentage of PyVRP. Fallbacks: without Advanced GNN, Graph ML ch. 5 and ch. 8; without Control, its ch. 3 as self-study.
Y2.5 Fraud alerts you can defend. Both routes, finance (anti-money laundering). Courses: Advanced GNN ch. 6–7, Geometric Data ch. 2–3, LLM ch. 7 and 9.3, Boosting ch. 1–4, Security ch. 8–9, MLOps ch. 5. Data: the Elliptic Bitcoin transaction graph, with Elliptic++ and the IBM synthetic AML data. Baseline: a tuned GBDT on transaction features. Success: recall at a fixed alert budget (for example 100 per time step) above the baseline, with a CI that excludes zero. Fallbacks: without Advanced GNN, Graph ML 5.6–5.8 (TGAT, TGN); without Security, ESP ch. 5 and Advanced GNN ch. 7.
Source map by period
Most briefs lean on the same few chapters, met in this order. The Period 2 themes are taught during the build phase, so read the chapter a brief names before the week that needs it.
| Period | Chapters | What the project takes from them |
|---|---|---|
| Year 1, Period 2 | Graph ML ch. 10; Text Mining & NLP ch. 6–7 | GraphRAG and the hybrid; relation extraction, transformers, a minimal RAG |
| Year 2, Period 1 (core) | LLM ch. 6, 7.3–7.4, ch. 9 | serving and its cost; RAG evaluation and GraphRAG; benchmarks, judges, hallucinations, attacks |
| Year 2, Period 1 (core) | Geometric Data ch. 2–3, ch. 6–7 | Laplacian spectra and descriptors; deep learning on surfaces; equivariance |
| Year 2, Period 2 (if taken) | MLOps ch. 1–5 | lineage, CI gates, serving SLOs, euros per million tokens, drift monitoring |
Data licence and consent checklist
Check these five points at W0, before any data enters the repository. The charter records the answers.
- Licence. Name the licence of each dataset and model. Several forbid commercial use, such as MIND (Microsoft Research licence, research use only) and ABO (CC BY-NC 4.0).
- Redistribution. Keep third-party data out of the repository unless its licence allows redistribution; commit the download script and a hash of the files instead.
- Consent. Collect signed consent before recording anyone’s voice or asking people to rate outputs (user sessions, reviewer panels). Say what is recorded, for what purpose and for how long.
- Retention and deletion. Set a deletion date for personal data, for example voice recordings deleted at the end of the study, and keep that data out of logs and training sets unless the consent covers them.
- Datasheet. Write a datasheet for every set the team builds (query sets, request sets, ratings), in the format of the ethics seminar.
Calendar, deliverables and rubric
The programme publishes no calendar, deliverable list or grading grid for this module. What follows is a proposal of these notes, shared with the Year 2 projects. At the usual 25–30 h per ECTS, 8 ECTS is about 200–240 h of work per student.
Calendar
| Week | When | Milestone | Evidence at the end |
|---|---|---|---|
| W0 | last weeks of Year 2 Period 1, part time | Two-page charter: one question, one metric and one baseline per work package; “done” as usage-based acceptance tests; a module map (2–3 chapters per work package); a compute and euro budget; the data checklist | charter reviewed by a tutor; a first baseline number |
| W1 | Year 2 Period 2 | Walking skeleton: the thinnest end-to-end version, with the evaluation harness | one end-to-end run on a small query set |
| W2 | Period 2 | First advanced work package | ablation over at least 5 seeds |
| W3 | Period 2 | Mid-term review: the charter re-read, the backlog re-prioritised | acceptance tests re-run; review notes |
| W4 | Period 2 | Red-team swap (trust work package) | attack success rate per family, before and after fixes |
| W5 | Period 2 | Full evaluation; model card, datasheet, guarantee sheet; ethics memo | evaluation report with confidence intervals |
| W6 | Period 2 | Paper and live demo; the defence follows | tagged release |
Deliverables
- Code: a repository runnable with one command, a tagged release, a deployed service or demo, and lineage for data, prompts, indexes and models (MLOps ch. 1).
- Paper: 6–8 pages in English, workshop style, plus appendices.
- French summary: one page for a non-specialist reader, the Languages hook.
- Documentation: a model card, a datasheet, the guarantee sheet of the trust work package, and the decision and technical-debt log.
- Defence: 20 min including the live demo, then 10 min of questions.
- Next steps: the case memo of the ethics seminar, written on this project, and a one-page open question for the final internship or a PhD / CIFRE proposal.
Rubric (proposal)
| Criterion | Points | Full marks when |
|---|---|---|
| Usage-based acceptance tests | 15 | tests fixed at W0, then passed or honestly failed |
| Core: Large Language Models and Geometric Data | 20 | the core methods are correct and beat or match the baseline for a stated reason |
| Advanced work packages | 10–15 | each advanced course used answers a question of the charter |
| Evaluation and statistics | 15 | at least 5 seeds, paired confidence intervals, a test set touched once (shared evaluation kit) |
| Trust work package and red-team swap | 15 | the four pieces of the trust work package, with adaptive attacks |
| Engineering and operations | 10 | serving SLOs, cost per request, CI gates |
| Reproducibility and decision log | 5 | every number regenerated with one command; decisions dated |
| Communication and ethics memo | 5–10 | paper, French summary, defence, case memo |
Two rows carry a range: choose their weights so that they add up to 20 points, and the total to 100. For example, 15 for the advanced work packages and 5 for communication, or 10 and 10 when the ethics memo is graded here. Each brief on the projects page adapts the rows to its own question.
Trust work package
Every Year 2 brief carries a mandatory trust work package (objective O8). It turns “is the system safe?” into evidence, built from exercises that already exist in the notes.
- Threat model. Before any attack, write the threat model in the format of Security & Robustness Definition 1.1: the attacker’s capabilities, the timing and the goal, plus the attack budget and what is out of scope.
- Attack suite. Follow the red-team protocol of ESP Ex. 11: 30 attacks, 10 of them adaptive (20 and 5 for a solo project), scored by attack success rate per family. LLM 9.4 is the mandatory-path source for jailbreaks and prompt injections.
- Guarantee sheet. State one uncertainty or privacy guarantee, such as a conformal coverage or a privacy budget, on one sheet in the format of Privacy & UQ Ex. 11: what is guaranteed, under which assumption, measured on which data, and what would break it.
- Regression tests. Every successful attack becomes a CI test, so that a later change cannot reopen it (MLOps ch. 2, if taken).
Red-team swap (W4). For one week, each team attacks another team’s system with a fixed budget, like the game day of MLOps Ex. 12. The defenders report their detection time and their fixes, and the attackers’ successful cases join the defenders’ CI suite.
Security & Robustness, ESP and Privacy & UQ are slot TBC. Privacy & UQ is announced for Period 1, but the other two for Period 2, after W0. Read the definition or exercise named above before the week that needs it, as self-study if its course is not offered or comes later: the threat model at W0, the attack suite and the guarantee sheet by W4.
Extrapolating Framing, subject bank, calendar, rubric and trust work package are proposals of these notes; brief and grading = official project setup.