This arXiv study finds a hyperbolic VAE retrieves cross-category prior art 52% more often than a matched Euclidean model on 45,000 US patents, though it agrees less with official classification codes.…

#PatentSearch #EmbeddingModels #HyperbolicML #UnsupervisedAI
https://arxiv.org/abs/2610.12997

New arXiv work analyzes how Riemannian gradient flow in hyperbolic space implicitly biases multiclass classification, linking a radial drift coefficient to the Busemann function and showing…

#MachineLearning #RepresentationLearning #HyperbolicML
https://arxiv.org/abs/2610.07131