News & feedback · 16 November
View your A2 marks & feedback Use your GitHub ID and the email entered in the registration sheet.
A3: create AND complete your group by Friday 20 November evening
Groups must contain three students. Create your group and make sure all three members join it by Friday 20 November, in the evening. An empty or incomplete group does not meet the deadline. On Saturday 21 November, in the morning, I will randomly assign students without a group and complete the remaining groups, with a 2-point penalty. Join A3 and check the requirements.
Juliet: start early — submit jobs now!
If you plan to run on Juliet, submit your first jobs as soon as possible. Test your environment, imports and data with a short job now; allow time for the queue and experiments. Do not wait until the final night.
From A2 to A3: investigate, do not just describe
Focus on one method, or two only if genuinely complementary. Explain why they complement each other. Formulate hypotheses about why something works or fails, then design controlled experiments to test those hypotheses. A better score alone is not an explanation.
Slides de Lucas Gnecco and A2 feedback slides are available in the A3 section.
A3 finals: 3 December, 14:15–17:30; 4 minutes per presentation.
Push slides.pdf and report.pdf by 2 December at 23:59;
code can be updated until 3 December at 05:00 (Paris time).
Report guidelines · A1 marks & feedback.
Find your A3 group and check the available places. Make sure your group has three members by Friday 20 November, in the evening, with no pair previously grouped together in A1 or A2.
The course
Work in groups of three on collaborative filtering, generative models and adversarial robustness. For each project, implement an approach, compare it against baselines, and explain what your experiments teach you in a short talk and a report.
Registration & groups — action required
Your TODO
- Complete your name and email in the registration sheet, next to your GitHub ID.
- A2: choose new teammates. No pair of students from an A1 group may work together again in A2. Check the group rule before joining a team.
- A2 presentation dates are available: check your group for 5 or 12 November, and check your membership and the no-repeated-pair rule. These lists are not the running order.
- A2 preliminary presentation order · 5 November
Each group has 5 minutes to present; be ready when the previous group finishes.
- A3: groups of three and 4-minute presentations. All groups present on 3 December, 14:15–17:30. Slides and report are due on 2 December at 23:59; code can be updated until 3 December at 05:00 (Paris time).
- A2 final presentation order · 12 November
Each group has 5 minutes to present; be ready when the previous group finishes.
Assessment & submissions
- One presentation per group, per project. A1 and A2: preliminary or final. A3: final only, for everyone. A2 allocation: 19 groups in preliminary and 20 in final. A1/A2 presentations last 5 minutes, with an alarm at 5 minutes and a further 2 minutes for questions and transition. A3 presentations last 4 minutes, with an alarm at 4 minutes. The A3 arrangements for questions and transitions will be announced separately.
- One final report per project: one group cover page + at most five pages of content. Optional appendices are not read.
- Code itself is not graded. Your report and talk must explain your methods, experimental protocol, results and limitations. The required code interface must work for automated tests.
- AI is allowed. You must understand, verify and be able to defend everything you submit. Interpret results rather than simply describing them. Expectations increase from A1 to A3.
Files and deadlines
Push slides.pdf and report.pdf to the root of your group
repository. Slides are due at 23:59 the day before your presentation; the final report
is due at 23:59 the day before the final session, for all groups. Code can be updated until 05:00 on
the final-session day. All times are Paris time.
| Project | Preliminary slides | Final slides & report | Final code |
|---|---|---|---|
| A1 | 14 Oct · 23:59 | 21 Oct · 23:59 | 22 Oct · 05:00 |
| A2 | 4 Nov · 23:59 | 11 Nov · 23:59 | 12 Nov · 05:00 |
| A3 | — | 2 Dec · 23:59 | 3 Dec · 05:00 |
Late submissions will not be accepted. Missing or incorrectly named files will be penalised. Check your pushed PDFs on GitHub before the deadline.
Daily testing
Your repository is pulled and tested each morning around 05:00. Results appear after the runs finish. The DataLab results links will be announced below. These tests provide feedback for your experiments.
Schedule
A1: 40 active groups · allocation updated on 15 October. Each presentation lasts 5 minutes, with an alarm at 5 minutes, followed by 2 minutes for questions and transition. Preliminary: 22 groups · 154 minutes (2 h 34). Final: 18 groups · 126 minutes (2 h 06). Group lists and presentation dates are below.
Sessions generally last three hours. The A1 preliminary session on 15 October uses an extended presentation window of 2 h 34. On 22 October, the last 40 minutes are reserved for the introduction to A2 and precision–recall. A2 group allocation remains provisional.
| Date | Session |
|---|---|
| Thu 1 Oct | DataLab introduction + A1 |
| Thu 8 Oct | Slurm introduction + technical support |
| Thu 15 Oct | A1 preliminary presentations |
| Thu 22 Oct | A1 final presentations + A2 / precision–recall introduction |
| Thu 29 Oct | No class |
| Thu 5 Nov | A2 preliminary presentations |
| Thu 12 Nov | A2 final presentations + A3 introduction |
| Fri 13 Nov | Guest lecture: Lucas |
| Thu 19 Nov | No class |
| Wed 25 Nov | No class |
| Thu 3 Dec | A3 final presentations — all groups · 14:15–17:30 |
Collaborative filtering
Predict missing ratings using matrix factorisation and alternative recommendation methods. Compare approaches and study hyperparameters, regularisation and computational cost.
--name.Preliminary · Thursday 15 October
A1 preliminary presentation order · 15 October
22 groups · 5 minutes of presentation per group. An alarm rings at 5 minutes. Allow a further 2 minutes for questions and transition.
Upload slides.pdf by 14 October at 23:59.
Groups · alphabetical list
- 1p2v
- ACA
- AcceleratedAerodynamicsYaris
- AKE
- Antoine
- BANs
- Blocus
- blue
- Gengis GAN
- Group
- Group 28
- Group 4
- Group 7
- HBA
- La piscine
- Le Groupe
- Le groupe du fond : LiS
- MassSurveillanceReadyFiltering
- ocaml
- OT dream team
- rEcommeNderS
- WorksOnMyMachine
Final · Thursday 22 October
A1 final presentation order · 22 October
18 groups · 5 minutes of presentation per group. An alarm rings at 5 minutes. Allow a further 2 minutes for questions and transition.
Upload slides.pdf by 21 October at
23:59.
Groups · alphabetical list
- altf4
- Breaking Bias
- Cold Start Club
- data-doodlers
- DauDau
- force-push
- Group 13
- Group 16
- Group 31
- Group 37
- LHOOQ
- ONN
- Rejoignez@Alexis@Hadrien
- SnapCat
- The Outliers
- TinderCEOs
- umap reborn
- UsualGreenSundayEnjoyer
The three volunteer groups are included in the preliminary session. The remaining preliminary groups were drawn at random. These alphabetical lists indicate your presentation date, not your speaking order. The preliminary running order remains available for reference. The final presentation order is now available above. Please be ready from the start of your session.
All groups: upload report.pdf by 21 October at 23:59,
including groups presenting on 15 October. Code remains modifiable until 22 October at
05:00.
Check the latest assignments on the Find your A1 group page. Contact Alexandre if your group is missing or incorrect.
Generative models
A2 groups: no repeated pairs
A2 groups created after 3 November will receive a grade penalty. Creating a group late does not exempt it from selection for preliminary presentations.
No two students who were in the same A1 group may be in the same A2 group. Every pair in your A2 group must be new. For example, if A, B and C worked together in A1, they must join three different A2 groups.
Study generation quality and diversity through precision and recall. Design controlled comparisons and explain the trade-offs you observe.
Train the MNIST GAN with the provided generator architecture kept fixed. The platform measures FID, precision and recall and displays generated-image galleries.
Preliminary · 5 November
A2 preliminary presentation order · 5 November
5 minutes of presentation per group. An alarm rings at 5 minutes. A further 2 minutes are reserved for questions and transition.
19 groups · 133 minutes including questions/transition.
Alphabetical list; not the running order. Running order drawn on the day.
- Bach enjoyers
- CirylGan
- CLM
- Consistent GANaches
- CreativeConcepts
- DauDauBis
- DLR
- GANgnam style
- Group 17
- Group 24
- Group 37
- Klick
- LaMasIA
- Latent Dwellers
- Madeleines
- Réveille-toi Canna
- Sbo3a
- The Ambassadors
- TP
Final · 12 November
A2 final presentation order · 12 November
5 minutes of presentation per group. An alarm rings at 5 minutes. A further 2 minutes are reserved for questions and transition.
20 groups · 140 minutes including questions/transition.
Alphabetical list; not the running order. Running order drawn on the day.
- 3StochasticParrots
- A2
- AKA
- AR
- blue
- Data Lab - 2
- ganberge
- GANsters
- Group 33
- HQHD
- jaipasdidéesdenomdegroupe
- Lamine Yamal
- MNISLOP
- NYBandits
- QOR
- Que la GAN
- RTFM
- SES_team
- Skill issue
- Tung Tung Tung Tung Tung GANur
A2 session allocation
TP and The Ambassadors volunteered for the preliminary session. The additional preliminary groups were selected at random on 31 October; the other groups present in the final session. The lists above give presentation dates, not the running order.
Slides are due at 23:59 on 4 November for preliminary groups and 23:59 on 11 November for final groups. The final report.pdf is due on 11 November at 23:59 for all groups. Code can be updated until 12 November at 05:00. All deadlines use Paris time.
Contact Alexandre if your membership or presentation date needs correction. New groups require an explicit session assignment; refreshing the group list does not redraw existing dates.
Adversarial robustness
A3: no pair of students may have worked together in A1 or A2. Re-form any conflicting group by Friday 20 November, in the evening. Check the pair audit.
Daily evaluations start on 13 November, around 05:00 Paris time.
Understand adversarial attacks and assess the robustness of classifiers. Support your conclusions with careful experiments and critical analysis.
Final presentations · 3 December
Work in groups of three. All groups present in the final session. There is no preliminary session for A3.
The final session is on Thursday 3 December, 14:15–17:30 (Paris time).
4 minutes of presentation per group. An alarm rings at 4 minutes: finish your talk when it rings. With 39 groups, the talks take 2 h 36; the remaining 39 minutes are shared between questions, transitions and a break. The detailed running order will be announced separately.
Push slides.pdf and report.pdf to the root of your group repository by
2 December at 23:59. Code can be updated until 3 December at 05:00.
All times are Paris time.
The report is limited to five pages of scientific content, plus a group cover page. Optional appendices are not read. Focus on a well-supported experimental study, explain your choices, analyse the results and limitations, and use fair comparisons.
Group list, order and speaking times: to be announced.
MesoNET & Slurm
Prepare your account before the practical introduction on 8 October.
- Request an account through the MesoNET account procedure, using eduGAIN and your institutional account. Mention IASD.
- Validate your email, wait for approval, then complete your MesoNET profile before emailing Alexandre.
- Email Alexandre with your name, MesoNET username and registration email to request access to the IASD project.
- Prepare your SSH public key. The course introduction contains the SSH setup instructions.
The example includes Python code, configurations, Slurm scripts, the interactive analysis notebook,
recorded results and the two trained best.pt models. Start with README.md.
You can explore the prepared results without retraining. Download CIFAR-10 separately to train;
the larger resume checkpoints and environments are not included.