Geometric Deep Learning — Cohort site

Neural networks that respect the symmetries of their input domain — translation, rotation, permutation, graph structure. The mathematical thesis that ties GNNs, equivariant networks, and AlphaFold together.

self-study Self-study reference — no active cohort

Neural networks that respect the symmetries of their input domain — translation, rotation, permutation, graph structure. The mathematical thesis that ties GNNs, equivariant networks, and AlphaFold together.

→ Weekly schedule EN notes (PDF) FR notes (PDF)
Level Graduate
Instructor Dr. Yaé Ulrich Gaba
Meeting pattern Tuesdays 14:00–16:00 lecture · Thursdays 14:00–15:00 paper discussion (Africa/Lagos UTC+1)

Prerequisites

Linear algebra and multivariable calculus. Basic deep learning (forward pass, backprop). Familiarity with PyTorch. No prior group theory required — week 2 builds it from scratch.

Grading

Four problem sets (40%) · three code labs (25%) · paper-discussion recitations (10%) · final project (25%).

Reading

Bronstein, Bruna, Cohen, Veličković, Geometric Deep Learning: Grids, Groups, Graphs, Geodesics, and Gauges (2021) — recommended companion.

What this site is and isn’t

The bilingual notes (linked above) are the reference text. This cohort site is the operational layer: every week page has the lecture topic, the readings to do beforehand, the problem set or code lab, and any paper discussion. The schedule and weeks are generated from a single data file (_data/apprentissage-geometrique.yml), so the same source drives the landing, the schedule, and every week page. If you are reading along without being in a cohort, the week pages still work as a self-study guide; the deliverables become optional, but the readings and lecture topics are the same.