Tag
This paper investigates the impact of using hyperbolic versus Euclidean latent geometry for tree-structured regularizers in prototype networks, showing that hyperbolic prototypes significantly improve local topology preservation in hierarchical classification on WikiArt.
A follow-up project visualizing GPT-2's 32,070 tokens as an interactive hyperbolic tree in a Poincaré ball, allowing users to fly through the embedding space.
This paper proposes Equivariant Poincaré ResNets, combining hyperbolic geometry with discrete symmetry groups to improve efficiency in learning visual representations by treating rotated features as symmetric rather than distinct hierarchical concepts.