Boot camp on invariances in graph learning

on YouTube ↗01h 29m

" summary

Equivariance work focuses on concrete constraints and recipes: use representation theory, irreducible representations and Schur’s lemma to force linear layers to mix only matching irreps; use tensor products and Clebsch–Gordan decompositions for equivariant nonlinearities; practical pipelines combine pairwise distances with message passing, gated scalars, or spherical harmonics; alternatives are augmentation or frame averaging (interpolating between single canonicalization and full augmentation) to trade compute for continuity; theoretical universality follows from group averaging, equivariant polynomials, or orbit separation (e.g. deepsets, 2‑WL for R^3 point clouds).

AI-written summary · about

" chapters (3)

1. 0:00Mathematics of equivariance
2. 27:46Geometric graph networks
3. 59:02Expressivity and universality

" related talks

• 01h 09mEquivariant RLMax Welling · Optimization Under Symmetry 2021
• 47mInvariance, Geometry and Deep Neural NetworksPavan Turaga · CVPR 2020
• 59mEquivariant Machine Learning Structured Like Classical PhysicsSoledad Villar · Optimization Under Symmetry 2021
• 54mLearning Uninformative RepresentationsRichard Zemel · Adversarial Approaches in Machine Learning 2022
• 30mPreconditioning In Sparse Linear Regression Using Graphical StructureFrederic Koehler · Algorithmic Aspects of Causal Inference 2022
• 51mRandom Projections for Probabilistic InferenceStefano Ermon · Uncertainty in Computation 2016
• 51mOptimization On Tensor Network VarietiesFulvio Gesmundo · Optimization Under Symmetry 2021
• 45mScale Learning and Reasoning Across Heterogeneous Gradients and SemanticsTian Li · Learning from Heterogeneous Sources 2026