[M2L 2025] 4.2 A Broad Introduction to Topological Deep Learning - Mathilde Papillon

Value of the Information & Strength of the Argument

The presentation provides a clear and accessible introduction to GNNs and TDL, building intuition from basic concepts to more advanced topics. The argumentation is logical: starting with the limitations of pairwise-only GNNs, she motivates the need for higher-order relations and introduces TDL as a solution. She effectively uses analogies (e.g., convolution on images) and visual examples to explain complex ideas. The value lies in its pedagogical approach, making it suitable for newcomers to the field. However, the talk is introductory and does not delve into mathematical details or empirical comparisons, limiting its depth for advanced practitioners.

Scientific Rigor, Source Quality, Title Accuracy

The talk references key works in the field, such as the GCN paper by Kipf & Welling and message passing neural networks, but does not provide specific citations or URLs. The content is consistent with established knowledge in the field. The title accurately reflects the content, which is a broad introduction to TDL. The talk is well-structured and scientifically sound, though it lacks explicit source citations and in-depth technical details. The Q&A session addresses relevant concerns, showing responsiveness to audience questions.

The title accurately reflects the content: a broad introduction to topological deep learning, starting with GNN basics.

8/10

Presentation by a researcher (Mathilde Papillon) at a summer school, covering established concepts (GNNs, TDL) with references to key papers (Kipf & Welling, message passing). The content is well-structured and accurate, but lacks in-depth mathematical derivations and external verification.

Markers derived by PSI from the transcript: the creator did not define chapters.

• 00:00Introduction and outline of the talk.
• 02:00Motivation for graph learning: social networks, molecules, Google Maps.
• 05:00Explanation of permutation invariance and equivariance.
• 08:00Introduction to adjacency matrix and basic graph convolution.
• 10:00Derivation of the mean pooling layer and GCN (Kipf & Welling).
• 12:00Introduction to message passing neural networks.
• 15:00Discussion of limitations of GNNs: over-smoothing and higher-order relations.
• 18:00Introduction to topological deep learning and higher-order domains.
• 21:00Explanation of simplicial complexes, cell complexes, hypergraphs, and combinatorial complexes.
• 24:00Definition of boundary, co-boundary, and adjacencies on higher-order domains.

• Semi-Supervised Classification with Graph Convolutional Networks — Referenced as the most famous GNN (Kipf & Welling).
• Neural Message Passing for Quantum Chemistry — Referenced as the message passing neural network paper.

• Graph Neural Networks: A Review of Methods and Applications — Provides a comprehensive overview of GNNs, consistent with the talk's introduction.
• A Practical Tutorial on Graph Neural Networks — Offers a practical introduction to GNNs, aligning with the talk's pedagogical approach.