{"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","summary":"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…","category":"research","comments":0,"disputed":false,"division":"math","mentions":0,"discussed":0,"published":"2026-02-26","reactions":{}}