#vector-bundles News & Analysis

Learning Tangent Bundles and Characteristic Classes with Autoencoder Atlases

Researchers introduce a theoretical framework connecting multi-chart autoencoders in manifold learning with classical vector bundle theory and characteristic classes. The approach treats collections of locally trained encoder-decoder pairs as learned atlases on manifolds, enabling computation of differential-topological invariants and providing algorithmic criteria for detecting orientability.