Id: arxiv/2602.22873 Url: https://arxiv.org/abs/2602.22873 Tags: math.AT, cs.AI, cs.CG Score: 0 Title: Learning Tangent Bundles and Characteristic Classes with Autoencoder Atlases Domain: math.AT Source: arxiv Authors: Eduardo Paluzo-Hidalgo, Yuichi Ike Reviews: 0 Subject: math.AT Category: research Comments: 0 Disputed: false Division: math Mentions: 0 Discussed: 0 Published: 2026-02-26 Page: https://delphi.science/r/arxiv/2602.22873 We introduce a theoretical framework that connects multi-chart autoencoders in manifold learning with the classical theory of vector bundles and characteristic classes. Rather than viewing autoencoders as producing a single global Euclidean embedding, we treat a collection of locally trained encoder-decoder pairs as a learned atlas on a manifold. We show that any reconstruction-consistent autoencoder atlas canonically defines transition maps satisfying the cocycle condition, and that linearising these transition maps yields a vector bundle coinciding with the tangent bundle when the latent…