HYDRA: Hyperbolic Dynamic Representation Architecture for Kolmogorov-Arnold Networks
HYDRA cuts KAN redundancy by sharing functional transformations inside a hyperbolic latent space.
The paper proposes a parameter-efficient version of Kolmogorov-Arnold Networks that maps vector inputs into the Poincare ball. It applies KAN-style updates in tangent space, then uses a low-rank prototype block to share transformations across hidden dimensions. The authors say radius control helps training stability by avoiding boundary saturation. Across eight benchmark datasets, they report competitive or better prediction with improved parameter efficiency and interpretability. ArXiv · AI/CL/LG's note
The paper proposes a parameter-efficient version of Kolmogorov-Arnold Networks that maps vector inputs into the Poincare ball. It applies KAN-style updates in tangent space, then uses a low-rank prototype block to share transformations across hidden dimensions. The authors say radius control helps training stability by avoiding boundary saturation. Across eight benchmark datasets, they report competitive or better prediction with improved parameter efficiency and interpretability. ArXiv · AI/CL/LG's note
score 4