An Idea From Physics Helps AI See in Higher Dimensions | Quanta Magazine

geometric-deep-learninggauge-equivarianceconvolutional-neural-networksphysics-ml

Abstraction: Gauge-equivariant CNNs extend deep learning to curved non-Euclidean surfaces

Key points:

Taco Cohen, Maurice Weiler, Berkay Kicanaoglu, and Max Welling (University of Amsterdam / Qualcomm AI Research) developed "gauge-equivariant CNNs" that can learn patterns on any geometric surface, not just 2D planes
Standard CNNs rely on translation equivariance; gauge CNNs generalize this to arbitrary curved manifolds by borrowing gauge equivariance from physics (general relativity, Standard Model)
On global climate data (spherical), gauge CNNs detected tropical cyclones with 97.9% accuracy vs. 74% for standard CNNs and 94% for a sphere-specific approach
Data efficiency is a key benefit: equipping networks with geometric symmetry assumptions reduces training data requirements dramatically (e.g., lung cancer detection with 1/10th the data)
The framework subsumes all prior geometric deep learning approaches (rotation equivariance, spherical CNNs, 3D pose recognition) as special cases
Applicable to 3D drone/autonomous vehicle vision, organ shape analysis, particle physics data (4D spacetime)

Connections: Qualcomm · University Of Amsterdam · Geometric Deep Learning · Convolutional Neural Networks · Gauge Equivariance